Keynote: Building Intelligent Systems on Real-time Data

[00:00:00] In a world without data streaming, you might be waiting a while for that rideshare. [00:00:07] And that flight the agent booked you on? [00:00:09] It left 20 minutes ago. [00:00:12] But don't worry. [00:00:13] Your bank will catch the fraudulent purchases on your account? [00:00:16] Eventually. [00:00:17] Maybe. [00:00:19] Try not to think about it? [00:00:21] At work, you'd still be getting 3 a.m. alerts when the infrastructure breaks. [00:00:25] And having to listen to your CEO talk about digital transformation in the morning. [00:00:30] Because all your data is frozen here. [00:00:32] Or at least it was. [00:00:35] That's the world without data streaming. [00:00:38] That's the world without you. [00:00:40] And you and you and all of us. [00:00:43] The pioneers who transformed the world we live in today with real-time data. [00:00:48] And who are already building the world we'll live in tomorrow. [00:00:51] Real-time AI agents. [00:00:53] Instant analytics. [00:00:54] Unprecedented customer experiences. [00:00:57] A world that's faster. [00:00:58] Smarter. [00:00:59] Frankly, it's just better. [00:01:02] That's the world we'll continue to build. [00:01:05] Together. [00:01:06] Welcome to Current 2025. [00:01:12] Please welcome CEO and co-founder of Confluent, Jay Kreps. [00:01:17] Hey, everyone. [00:01:29] Welcome to Current. [00:01:31] I'm really excited about this. [00:01:33] This is our first time in New Orleans. [00:01:36] And it's a really awesome city to host this. [00:01:39] I hope everybody got a chance to get out there and see a little bit of the city yesterday. [00:01:44] If not, hopefully in the next few days. [00:01:47] We're going to do all the fun stuff. [00:01:50] We'll have some deep dives in Kafka, in Flink. [00:01:53] We'll have some great product announcements. [00:01:56] Really, this is one of the few times you can get these communities together to talk about what they're working on, what they're building. [00:02:03] Always the sessions are great, but the hallway conversations are even better. [00:02:08] And so I think we have a great lineup for today and tomorrow. [00:02:12] I'm going to be talking today about something that I think is on all of our minds, which is the rise of AI. [00:02:19] And how do we build intelligent systems? [00:02:22] How do we connect those into the software states that exist in our organizations today? [00:02:28] I'm going to connect that into the world of data streaming and real-time data and agents that take action on that. [00:02:36] But to set this up, I want to start and just talk a little bit about where we're at right now. [00:02:40] I think there's a sense in which we're kind of moving out of a world of business intelligence and into a world of AI. [00:02:49] And what do I mean by that? [00:02:51] I mean, if you ask an organization, what does it mean to be data-driven, the answers to that question would have changed. [00:02:59] If you asked that question 10 years ago, I think it would have been about, hey, how can we get all the data into our lake house, into our warehouse? [00:03:08] How can we be really smart in the reporting and analytics? [00:03:13] How can we arm our data scientists to address the right questions for our executives? [00:03:18] That's what it would mean to have good business intelligence. [00:03:21] That's what it would have meant to be data-driven at that time. [00:03:24] But if we think about what it means today, that's changing. [00:03:27] And it's changing in a way where it's no longer about just deriving insights from data. [00:03:33] It's about taking action. [00:03:35] It's about being able to build systems that actually act on behalf of the business in software, [00:03:41] that kind of close the loop in what we're doing. [00:03:44] And that's the age that we're moving into. [00:03:47] And in a sense, this is nothing new. [00:03:49] We think about the software systems that we've been building probably for the last five years. [00:03:53] A lot of them are doing this in some sense, right? [00:03:55] They're trying to be smarter. [00:03:57] They're trying to harness data to take action on behalf of the customer. [00:04:00] They're trying to personalize things and make better customer experiences. [00:04:04] So there's a sense in which this is not different. [00:04:07] But there's a sense, I think, in which it's really different, right? [00:04:10] Ultimately, AI systems are a really different world. [00:04:13] And it's easy enough to use Claude or ChatGPT and ask it some questions and get an answer. [00:04:21] It's actually a lot harder to build something that's a concoction of traditional code along [00:04:27] with AI that actually takes over some business process that automates some part of a big at-scale business and does this in a way that's high quality enough that it actually works, that it can actually go to production and actually fulfill its goal. [00:04:44] And part of the reason this is so hard is the paradigm is actually completely different, right? [00:04:48] If we think about how things have changed, the biggest difference is that traditional software is ultimately a bunch of hard-coded rules. [00:04:57] You know, it's a pile of logic and we've gotten incredibly good at crafting this logic, packaging it up, abstracting it away, coming up with modules and systems of isolating parts of this that we can support these fantastically complicated piles of business logic. [00:05:16] But at the end of the day, when you think about AI and AI-based systems, it's really not that. [00:05:22] You're not programming it in the same way. [00:05:24] You're not specifying it in the same way. [00:05:26] It's a little bit more like you're guiding it with data. [00:05:29] You're pushing it in a certain direction. [00:05:32] But ultimately, it's a probabilistic system, right? [00:05:35] The output is not precisely specified, or at least the output that we're looking for. [00:05:41] And, you know, this shows up in a really fundamental way when you think about how you develop and validate these systems. [00:05:48] In traditional software, you can reason about the correctness of your application without any reference to the underlying data. [00:05:57] You know, you can write a set of unit tests that cover the different cases. [00:06:00] If that unit test coverage is good enough, you expect 100% to pass, even though maybe all you're plugging in is just lorem ipsum, just some fake data that you're validating on. [00:06:11] And this is really different from an AI system. [00:06:14] At the end of the day, if you're building an AI system that's going to support your customers, there's no sense in which you can say that that system works if you haven't tried it on real customer issues with the real data that would have been available to a real support person who would have helped that customer. [00:06:30] If you haven't tried that, there's no amount of looking at your logic or the workflow or specifications or anything like that that would say that the system actually does what it's supposed to do. [00:06:41] And so fundamentally, the world of evaluation goes from being primarily about checking logic to now checking this combination of model, logic, and data all together in real interaction. [00:06:53] And checking it against a metric, which is not exact, it's not going to be 100% perfect, you're going to have a set of evals that tell you you're better, you're 90%, you're 95%, you're improving. [00:07:06] You know, it's much more a statistical notion of correctness. [00:07:09] And really importantly, your development cycle is now very tied to the real data that you use. [00:07:16] In some sense, in traditional software, the business logic is king. [00:07:20] That's really everything, right? [00:07:22] The data is just something that lives in production that you're going to go use. [00:07:26] In these AI systems, ultimately, the data, and particularly the context data, is the fundamental thing. [00:07:32] That's how you're guiding the model. [00:07:34] That's what it has to go off of. [00:07:36] And making that really good is your ticket to making this actually effective. [00:07:42] So how can you build effective systems in this world? [00:07:45] What are the important problems that we have to solve? [00:07:48] If this is different in some way, how do we have to adapt what we're doing? [00:07:53] Well, I'll talk about some aspects of this. [00:07:55] So, you know, I mentioned that this was ultimately about context data. [00:07:59] And this is, you know, this is a really important part of the problem. [00:08:01] If you think about the success of these systems, the two things that are going to matter in terms of the quality of the output are the model and the context data that it has, right? [00:08:11] And I'll tell you, one of those things you have a lot of control over, right? [00:08:15] The models are being developed by a relatively small set of companies that are taking that forward. [00:08:20] You can always upgrade to the next best one, but your ability to advance that state-of-the-art is not that high. [00:08:26] The context data that is spread all across the organization that you want to harness for these problems, that's where you can iteratively make things better. [00:08:34] That's how you can take a system from 90% of good enough to 100% of good enough. [00:08:40] That's how you can actually take a system from demo to productions by getting that right. [00:08:44] So getting really good at this is key to this area. [00:08:48] And when you ask how can you really harness context data, I think people start with a basic intuition that kind of makes sense. [00:08:55] When we think about context data for models, there's something called the model context protocol. [00:09:00] And if we hooked up the systems we had to an agent, it should be able to then get the context data that it needs. [00:09:06] So in some sense, the naive approach here would be, let's go put MCP in front of all the systems we have. [00:09:11] We'll give the agent kind of the login to all of our databases and SAS systems. [00:09:17] And we'll say, okay, you figure it out. [00:09:19] You know, go make sense of this mess. [00:09:21] And anybody who's tried this knows that it doesn't quite work. [00:09:25] That's not actually quite how it works. [00:09:27] That's not going to get you the result that you're looking for. [00:09:31] And the problem here isn't MCP. [00:09:32] MCP is great. [00:09:35] But MCP is just a way of accessing a system. [00:09:39] It doesn't actually fix what's in the system. [00:09:41] It doesn't actually curate that in a way that's useful for decision making. [00:09:46] And so ultimately, the data that we're accessing is the thing that's a problem. [00:09:50] So what goes wrong if you try and do this? [00:09:53] Well, there's a set of challenges. [00:09:55] I mean, the obvious one is you're connecting this new agent into a bunch of systems where you're going to generate operational load and workloads it wasn't really built for. [00:10:04] But that's not even the real problem. [00:10:05] The bigger problem is about access control. [00:10:08] Ultimately, if I'm serving customer A, it can't be the case that there's any probability, no matter how small, that I'm going to leak data from customer B or that it's going to have access to things that it shouldn't. [00:10:19] And ultimately, a lot of the data that we have in the existing applications and systems is hidden behind bad APIs or it's very much tied to obscure inner details of our applications. [00:10:36] You know, what does it mean when this column has code four? [00:10:41] I don't know, but you probably have to go check some enum in the code base to really make any sense of that. [00:10:47] And it's not like models intuitively understand the meaning of four any more than humans do. [00:10:51] Right? [00:10:51] At the end of the day, you have to have the full context to be able to make a decision. [00:10:55] So when you think about harnessing this data, you're ultimately going to have to build a useful data set. [00:11:01] You're going to have some kind of data pipeline that takes data out of the production environments, does some kind of transformation on it, builds it into the appropriate context that an agent may need to act, and serves it back up on demand. [00:11:16] You know, this is ultimately the pattern that's going to be really important. [00:11:19] There may be some things where you can just hook it up with MCP, but the most common thing you're going to have is working on this context data. [00:11:26] And this is commonly called context engineering, this idea of refining, processing data from different systems, preparing it for use in an agent. [00:11:36] So how can we do this? [00:11:37] How can we make this really effective? [00:11:40] If this is the key for us to be successful with AI is really mastering context data, how can we do it well? [00:11:46] Well, the obvious place to start would be how do we build these data pipelines elsewhere in the organization? [00:11:53] And the answer there is commonly batch processing. [00:11:56] A really simple way to start would be go to the data warehouse or lake house, there's a bunch of data sitting there, build out a set of processing steps, and prepare that data for access by the AI. [00:12:10] And, you know, we're actually quite successful at this, you know, as part of our analytics flows. [00:12:16] You know, a fair amount of rich processing goes into preparing data for analysis. [00:12:20] So we might end up with an architecture kind of like this, you know, something in your lake house or warehouse. [00:12:26] We're kind of running a set of iterative processing steps. [00:12:30] And then the output of that, you might upload to some kind of live-serving database where it could serve live queries from the agent. [00:12:38] So what's the problem with this? [00:12:41] Well, you know, first of all, there's some good things. [00:12:44] I mean, we're actually quite agile at working with batch data. [00:12:47] We can be quite iterative. [00:12:48] So in a sense, we've solved this problem of being able to grab new ingredients, refine them, make them better. [00:12:54] That development workflow is actually pretty good. [00:12:57] But when it comes to taking it to production, it gets a little bit harder. [00:13:01] You know, first of all, we have this somewhat tenuous integration with the live-serving system where we're uploading big batches of changes all at once and hopefully not knocking that system over. [00:13:10] And anybody who's built that kind of reprocessing and indexing pipeline knows that can be a little bit squirrely. [00:13:18] But that's not even the big problem, right? [00:13:20] The big problem is ultimately this data is getting processed in batch. [00:13:24] You know, best case, we're talking maybe four hours later, we know what the state of things are. [00:13:29] More likely, it's tomorrow. [00:13:32] And, you know, for feeding context to a system that's going to be making real-time decisions, it's going to be interacting as part of the business, this is a complete non-starter. [00:13:42] You're going to get totally wrong answers. [00:13:45] You know, you could use the analogy of, you know, if you were going to cross the street, a busy street, would you be willing to do that if all you had access to was a photo of where the cars were yesterday? [00:13:59] And the answer is no. [00:14:00] That would be a very dangerous proposition. [00:14:03] But ultimately, if you have some software system that's going to be interacting with the real state of your business right now, then doing that off some snapshot of data is not going to be good enough for hardly any use cases that matter, where the consequences are important. [00:14:17] So you're ultimately going to have to move this into something that is real-time, something where the context data is in sync. [00:14:24] So how can you do it? [00:14:26] Well, we're here at Current, so, you know, part of the answer is going to be streaming data, right? [00:14:30] There's no surprise there. [00:14:32] But it's not as easy as just, okay, change it all to stream processing. [00:14:36] There's some things you have to do right to make this productive, right? [00:14:39] And so the architecture can be something like this, where you have streaming data capture, you know, something like Kafka or Kafka Connect, stream processing, something like Flink, some way of serving this up. [00:14:50] But there's a few things we're going to have to do well to make this actually productive. [00:14:55] So first of all, I talked about the fact that iterating on this data was really key to our productivity, that we had to have a whole evaluation loop around it. [00:15:05] And I said that one of the things that was really good in the lake house was that this batch processing was really productive for development. [00:15:14] I could actually try different ingredients, create a data set, throw it away, try it a different way. [00:15:20] You know, I could actually iterate really effectively. [00:15:23] And so it's going to be really important that we preserve that. [00:15:26] And, you know, in order to do that, you need to have a conception of this stream processing that's broader than many people often have. [00:15:36] It's not going to be enough to just have something that you can string together that processes the data of right now. [00:15:42] It's going to have to be able to process the historical data and the real-time data together. [00:15:48] I'm going to need something that's, you know, not a specialized system, but kind of a generalization of batch processing, where I can build my program and have it run offline against a test data set, evaluate that, and then have it run continuously as it goes to production. [00:16:05] So what do I need to do to accomplish this? [00:16:13] Well, this is something we've been working on at Confluent. [00:16:16] I think there's two critical ingredients. [00:16:17] You know, first, at the data layer, I have to unify batch and streaming. [00:16:24] And we've been doing this with what we call table flow. [00:16:28] And this is a way of taking the real-time streams of data that are coming in Kafka and be able to continuously populate and represent that as open data tables in cloud object storage using standards like Iceberg or Delta. [00:16:44] So that you have a historical data set as of a point in time, as well as the real-time flow of changes. [00:16:50] And these two are not just two distinct systems, but are actually linked. [00:16:54] You understand the point in time across these systems. [00:16:58] You can process seamlessly between the two things. [00:17:01] When you've done this, you now have the ability to build systems that process history and then go on and start processing the real-time stream that are, you know, effective in batch but also effective in real-time. [00:17:14] And on top of that, you have to actually do this in the processing layer. [00:17:17] And this is something that Flink brings, this ability to build a program, you know, with a particular set of logic in SQL or a programming language, run that offline as a batch process, be able to iterate on that, try out different things to get it right, and then have it run in a streaming fashion. [00:17:34] And be able to go back and do that over and over and over again. [00:17:37] Because that's ultimately going to be our development cycle. [00:17:40] That's going to be the iteration loop by which we make our system better. [00:17:44] So if we put these two things together, we can have effective streaming pipelines that actually make this, you know, easy to use in development in the way that our batch system was. [00:17:57] Able to produce real-time data, which is going to be essential for agents. [00:18:01] So we've really made a substantial step forward on the first part of this data supply chain. [00:18:07] But what about the last step? [00:18:09] So I said it's not that easy to strap the real-time serving onto the data lake. [00:18:13] Is it any easier if you're strapping it onto a stream processing system? [00:18:19] The answer is not really, right? [00:18:21] We're going to have a continuous flow of updates coming out of our streaming. [00:18:24] But every time we go back and reprocess things to improve our context data, we're going to have a big batch load. [00:18:30] It's going to be relatively hard to keep these two things in sync. [00:18:33] It's going to be hard to avoid disrupting these real-time queries that are looking up the right bits of context at the right time. [00:18:40] And we're going to have a set of problems, right? [00:18:42] There's going to be operational problems about the load that we're creating. [00:18:46] There's going to be data model problems of just matching thing A to thing B, carrying the security model through. [00:18:51] We're ultimately balancing across two distinct systems that we have to somehow keep in sync. [00:18:57] And very often these live-serving systems are just not built to take the volume of batch updates that come from real reprocessing of data. [00:19:07] And they're often impacted by the real-time load of streaming updates. [00:19:12] So how can we make this easier? [00:19:14] Ideally, we shouldn't have to think about this last step at all. [00:19:16] It should be just like an implementation detail that as we've created this new dataset, it's somehow ready to serve. [00:19:22] How can we accomplish that? [00:19:25] Well, we've effectively accomplished something like this with TableFlow, where you have a flow of real-time data from Kafka. [00:19:33] You may have processing on that in Flink. [00:19:35] And then it lands in a table that's structured where you could query it, do things with it. [00:19:42] What's the problem with these TableFlow tables that are in Iceberg and Delta? [00:19:46] Well, there's only one problem, which is they're slow. [00:19:49] They're pretty slow. [00:19:50] They're high throughput. [00:19:51] They're scalable. [00:19:52] But they're a little bit slow. [00:19:54] And so how can we speed this up? [00:19:55] If we could somehow speed this up, we could actually support fast queries off this materialized dataset. [00:20:02] And we could do that in a way that was well-integrated. [00:20:05] And, in fact, this is very possible, right? [00:20:07] Rather than strapping on a completely separate database, we could, you know, maintain a fast cache of this data. [00:20:15] And instead of just waiting until it flows out into the Parquet files, we could keep an optimized index that's kept in sync. [00:20:22] And we could have a fast path for queries that avoids, you know, the full processing runtime of Flink [00:20:28] and actually does, you know, fast analytics or lookups against this dataset. [00:20:34] And if we did this, we could have access for things that were indexed this way with milliseconds instead of multiple seconds. [00:20:41] And that would allow us to have, you know, two representations of this data without all the overhead of having to load back and forth [00:20:48] and manage these processes between two distinct systems. [00:20:52] And, you know, this concept of building materialized views in stream processing is not a new one. [00:20:58] But this makes it, I think, a lot easier to do. [00:21:01] And so we're excited today to be announcing early access to our real-time context engine. [00:21:07] And what this does is it takes the streams of data in Kafka, allows you to process and reprocess them, [00:21:14] and then provides access to the resulting data in a fast, low-latency model, which is exposed via MCP, [00:21:22] and which can be plugged into any agent or AI system built in any framework you like. [00:21:28] Like anything that works with MCP will plug into this data and have access to it. [00:21:32] And you can iterate on that data, evolve it, make it better, and keep it in sync over on the streaming side and serve it up in real time. [00:21:41] That's right. [00:21:42] That's right. [00:21:46] And so, you know, if you're interested in this, this is in early access. [00:21:50] There's a QR code. [00:21:51] You can read more about it. [00:21:53] I think, ultimately, this is a very powerful pattern. [00:21:56] This ability to do, you know, streaming data capture, stream processing, create a new data set, [00:22:03] be able to iterate on that in batch, and then take it out to production and have that data set evolve in real time, [00:22:09] be able to serve real-time queries against it. [00:22:11] It's useful for these workloads around AI and AI agents. [00:22:17] But if you have the query capabilities, it's actually useful for a lot of use cases. [00:22:22] Building and maintaining these kind of materialized fuses is useful for analytics. [00:22:26] It's useful for populating derived caches. [00:22:28] It's useful for observability. [00:22:31] We're just getting started. [00:22:32] You know, what we have right now is pretty simple queries with good performance. [00:22:36] We would love to expand that to richer sets of queries with fantastic performance across a wide range of use cases. [00:22:43] And this is something that we're actively working on. [00:22:46] So this is a little bit about our steps on, you know, how to build context data for agents. [00:22:54] But there's another problem we have to solve, which is how do we actually build the agents themselves. [00:23:00] We talked a little bit about this integration of batch and streaming that, you know, in this development model for AI, [00:23:06] I'm kind of working in batch offline until I get something that passes all my evals and seems good enough. [00:23:12] I'm kind of deploying it into production where it's going to run continuously and I want to have similar evaluation of performance coming off of it. [00:23:21] I need to do that for my context data, but ideally I would like to do it for the whole system, the data plus the agent plus everything together. [00:23:29] So how can I do that? [00:23:31] Well, this is where it gets more complicated because when people say agent, there's a bunch of things that they mean. [00:23:36] I mean, sometimes they don't even know what they mean. [00:23:38] But, you know, some people are talking about something that is effectively clicking on things on a computer. [00:23:44] Some people are talking about a chat bot that maybe has some background, you know, actions it can take. [00:23:51] I'm going to be talking about, you know, actually automating business tasks, taking some workflow that happens in a company and automating it, you know, with AI. [00:24:02] So this is kind of the definition that I'm focused on when we talk about this next section. [00:24:08] which is building streaming agents. [00:24:10] So agents that tap into the events that are happening in the business and react to what's there. [00:24:17] So, you know, this could be something where a sale occurs and there's a fulfillment process that the agent should take on and orchestrate parts of. [00:24:24] It could be something where an insurance claim is filed and part of that processing flow has steps where the AI is going to take over aspects of that. [00:24:32] A support ticket is lodged. [00:24:34] You want to actually process some of that. [00:24:36] So, you know, this is building agents that kind of work in the background and take over some chunk of business activity. [00:24:42] And to do this, we can use, you know, very much the same thing I described for context data. [00:24:47] The ability to plug something into the streaming model, run it on historical data sets until we get it right, launch it into a streaming world where it runs continuously. [00:24:59] And the architecture for this is actually quite simple, right? [00:25:02] You have an agent built in Flink. [00:25:05] It has access to an LLM. [00:25:08] The kind of resulting logic is going to be in your Flink code. [00:25:11] The LLM is going to be actually orchestrating a lot of the decisions. [00:25:16] The events coming in, you can tap into and respond to. [00:25:20] The actions going out are going to be another stream of events of what you're deciding. [00:25:26] This is actually a really powerful model. [00:25:28] One of the reasons it's powerful is because I can run it offline and have it not just, you know, make side effects in production, but actually produce this output stream of events. [00:25:39] And I can evaluate that like a data set. [00:25:41] It makes it really easy to score it. [00:25:44] And so this idea of streaming agents, we've written more about this. [00:25:47] This is something we've been developing with customers. [00:25:49] There's a blog post for this as well. [00:25:51] This is out in Open Preview. [00:25:53] You can use this in Confluent Cloud in our Flink service. [00:25:57] Check it out, and you can see more about that. [00:26:00] So I've talked about context data for agents. [00:26:03] I've talked about building these streaming agents. [00:26:06] There's one more thing, which, you know, for all this talk of AI, you know, sometimes we forget about the good old-fashioned machine learning. [00:26:15] And you might think, well, that stuff's like so old it's been around for, you know, multiple years. [00:26:22] But sometimes the old stuff is actually pretty good. [00:26:25] And, you know, it's true that these older machine learning algorithms, they don't have the power or the generality that modern foundation models have. [00:26:35] But they do have a couple of advantages. [00:26:38] One, they're really fast. [00:26:40] Two, they're comparatively extremely cheap to operate. [00:26:43] And both of those things are good, especially if you're dealing with high-volume data. [00:26:49] And there's a lot of areas which are just a little bit out of reach on the cost frontier with more agentic AI foundation models. [00:26:58] And so this same model that I described for plugging LLMs into streaming agents, this works with traditional ML algorithms as well. [00:27:09] And we're adding some of these. [00:27:10] We've started with one of the things that makes the most sense for streaming data, which is anomaly detection, forecasting, fraud detection, anything where you're looking at a pattern and looking for any perturbations or abnormalities. [00:27:27] This is something that, you know, effectively all our customers have in their problem set. [00:27:31] You know, detecting network outages, you know, monitoring business activities, looking at purchasing behavior, return behavior, stock trading systems. [00:27:41] And so there's a set of, you know, auto-fitting ARIMA models that we've built into Flink. [00:27:46] They're incredibly easy to use. [00:27:48] You don't need a team of data scientists to fit it. [00:27:50] It will just adjust to the stream that it sees and tell you what's out of the normal range. [00:27:55] What is the normal range? [00:27:56] What would you expect to see next? [00:27:58] You can plug these directly into your systems. [00:28:02] And for those interested in this, there's another QR code. [00:28:05] You can check this out as well. [00:28:07] So I've talked about three capabilities here. [00:28:10] You know, the context data, the streaming agents, traditional machine learning models. [00:28:15] To give us an easy way to talk about it, we wrapped all these together and gave it a name. [00:28:19] We're calling this Confluent Intelligence. [00:28:21] And these are three tracks that we're doing a lot of development on. [00:28:25] Thank you. [00:28:26] Thank you. [00:28:28] And so we're really excited about this area. [00:28:34] Each of these pieces of functionality kind of stand alone and can be used independently. [00:28:39] You know, you can use this context engine to build derived data sets to feed any model you like in any framework you like. [00:28:48] You can also use it to feed context data into streaming agents if you like that run inside of Flink. [00:28:55] You can also hook this together with these machine learning functions. [00:28:58] So I'll show you a little bit of how this fits together. [00:29:01] This is a common pattern we've seen with customers where you can actually link these things. [00:29:06] You have some stream of business events coming in. [00:29:09] You can detect anomalies on this. [00:29:12] So, you know, maybe this is some operational aspect of the business. [00:29:15] I want to see if it's working well. [00:29:16] If I see something unusual, then I want to say, aha, I don't recognize that. [00:29:21] That looks wrong. [00:29:22] But when I do that, usually what happens, it's going to go to some human and they get a notification to go look into it. [00:29:28] So wouldn't it be more helpful if I had the agent do that first passive work first? [00:29:33] Take the anomaly, dive in, try and figure out what's going on, and give kind of a first draft diagnosis, maybe a proposed remediation, or maybe just fix it if it can. [00:29:43] And so you can hook that up as the next stage in this pipeline, you know, keying off of the anomaly events. [00:29:51] And, of course, if I want to provide that agent that's doing this analysis, some kind of context data, I could feed it with this context engine. [00:30:01] So you can actually use all of these parts together as well as separately. [00:30:05] And so, you know, I think this is a really exciting area. [00:30:08] I think ultimately, if we think about our industry, when we think about using data, harnessing data, what are the systems and foundations we need for that? [00:30:17] I do think we're kind of moving out of a world where the most sophisticated use of data was about business intelligence. [00:30:25] It was about insights. [00:30:26] It was about reporting. [00:30:27] It was about analysis. [00:30:28] We're moving into a world where the most sophisticated use of data is about taking action. [00:30:33] It's about the things that happen in your production applications that happen in that world. [00:30:38] And to harness data in that world, you ultimately need a real-time streaming foundation. [00:30:44] And so we're really excited about what that means for the world of streaming and really excited about what this is going to open up, you know, in all of our customers and all of their use cases. [00:30:53] And to dive a little bit deeper into this, I'm excited to welcome a really important partner of ours from Anthropic. [00:31:03] So Anthropic, as many know, is a company really leading the way on developing the most cutting-edge foundation models. [00:31:12] Many of you have probably used their product as part of your coding workflows. [00:31:17] If you've used Cloud Code, you know, fantastic work that they've been doing. [00:31:22] And they've been doing it, I think, in a very safe and responsible way. [00:31:25] And so I'm really excited to welcome to the stage for a conversation on these topics Sean Falconer, our product leader for AI, and Rachel Lowe, the head of Applied AI from Anthropic. [00:31:37] So Sean and Rachel, welcome. [00:31:40] Really looking forward to what you have to say. [00:31:42] And thank you all very much. [00:31:44] Thank you all. [00:32:14] Well, Rachel, welcome to New Orleans. [00:32:28] Welcome to Current. [00:32:29] Thank you so much for being here. [00:32:30] Thank you for having me. [00:32:32] So there is a ton of big questions in the world of AI right now. [00:32:37] A lot of unknowns might say, you know, when is AGI coming? [00:32:42] When are we going to have Transformers? [00:32:44] And I don't mean Transformer models. [00:32:46] I'm talking about, like, Optimus Prime. [00:32:48] Yes, Bumblebee. [00:32:48] Turning my car into a robot. [00:32:50] Yeah, I like Bumblebee, too. [00:32:51] Yes. [00:32:52] So big questions. [00:32:53] So I've been thinking a lot about, you're trapped on stage with me. [00:32:56] What is the big question I can ask you? [00:32:58] And I think the thing that I settled on, which not only impacts me, but impacts everybody in this room, you've been to New Orleans a number of times. [00:33:07] I have. [00:33:08] Where do I go to find the best beignets? [00:33:11] That is a hard-hitting question. [00:33:13] I don't want to polarize this audience here, as I also personally have my own affinities. [00:33:21] And so I feel like we should ask someone that's much more neutral. [00:33:24] So why don't we ask Claude? [00:33:31] Okay, so let's see. [00:33:34] I love that it's already pulling up the classic, Café du Monde, we all love. [00:33:38] I've got a lot of local recommendations here, very exciting lists. [00:33:43] But the thing is, I need this tomorrow morning with my coffee at 8 a.m. [00:33:47] I don't have time today, unfortunately. [00:33:49] And so let's see if it is able to give me a slightly better recommendation based off of these parameters. [00:33:55] I, of course, know that I cannot leave New Orleans without a beignet. [00:34:00] Oh, a lot of specificity, some timing here. [00:34:04] All right, so I think you all will see me at Canal Street, Café Bendier, tomorrow morning. [00:34:10] Well, amazing. [00:34:10] Hopefully that helped everyone here. [00:34:12] But as smart as Claude is, imagine if it actually was sort of, knew what was going on in the business right now. [00:34:19] You know, what are the, what's the queue look like? [00:34:22] What are the inventory levels at the particular cafe? [00:34:25] So, you know, show up and, oh my God, the beignets are gone. [00:34:28] No one wants that. [00:34:28] Yeah, exactly. [00:34:29] So if we could connect actually Claude to sort of having awareness about the moment, and we'd be able to operate it sort of with the same awareness that the business would be able to operate in. [00:34:39] Wouldn't that be amazing? [00:34:41] Pretty powerful. [00:34:41] So, Jay talked a lot about this idea of intelligent systems, you know, not only learning from history, but being able to sort of act in the moment. [00:34:53] And in your customer conversations, like how is this idea kind of showing up with them and how are they thinking about navigating, you know, maybe the next year or so? [00:35:03] It's a really important question. [00:35:05] I think it's actually a question that we get from our customers quite a bit. [00:35:08] I think before I look forward in the next year of what's going to happen, it's important to also recognize where we are right now. [00:35:15] I think if we just even look at the conversations we were having around AI a year ago, it was really the chat bot era, I like to call it. [00:35:22] And even in the last six months, I think we were really focused on just this very linear input to output. [00:35:28] Yes, there is a bit of a LLM feedback loop there, but it was very, it was single step. [00:35:33] And we need to see and recognize that models have significantly evolved. [00:35:39] Even in just the last couple of months, we recently released Sonnet 4.5, Haiku 4.5 for the faster, more like daily tasks. [00:35:46] And I think that that's where it's become incredibly powerful to be able to build more multi-layer agentic architecture. [00:35:53] And I think that's really what we're having that conversation with more and more enterprises, where they're no longer thinking about it as like a single experiment. [00:36:02] And they're not thinking about it in more of like a sandbox within a portion of their company. [00:36:06] They're really looking across all of their different lines of businesses and thinking about where can I embed AI in here so this becomes a more agentic workflow. [00:36:15] And so that all the different teams that I have within my company are more expansive. [00:36:20] And I think that's a significant shift. [00:36:22] That is really exciting. [00:36:23] Jay is talking about exciting that we're talking about right now. [00:36:26] I mean, I think that that'll significantly help us think more about the future and power of AI as being tied to context. [00:36:35] And that is going to be very critical across the different lines of businesses that companies need. [00:36:40] Yeah, absolutely. [00:36:41] I mean, I think context might be the word of the day here. [00:36:43] You know, context is king. [00:36:45] It really seems to be the key for productionizing these AI systems. [00:36:49] So I guess, like, what does that actually mean in practice? [00:36:53] What are the challenges that these, you know, your customers face with actually bringing their enterprise data to power a system like Cloud? [00:37:01] Yeah, absolutely. [00:37:02] Yeah, absolutely. [00:37:02] There, first of all, there is the first step of recognizing that there is this shift. [00:37:07] I think, like, again, moving out of the areas of the chatbot, we really want to move to these multilayer agentic architectures that I keep talking about. [00:37:14] But what does that actually mean? [00:37:15] It's essentially thinking through a system that can reason, plan, and act so that we have this team of harmonious agents that are working together. [00:37:23] But again, like, when I talk to companies and kind of prompt them to think about what context they have, they think of that more as just, like, context engineering. [00:37:32] But it's so much more than that. [00:37:33] I think it's really thinking about the overall business-specific data that they have and not just thinking about it in terms of, like, what do they assume is data, oftentimes in more structured places, but thinking through, like, what is critical to my particular business from both a structured and a non-structured standpoint. [00:37:52] And then I think the second piece really is that it's easier than ever to build an agent. [00:37:58] I think Jay made that very, very clear just now. [00:38:01] For us at Anthropic, we're leveraging a number of different pieces to be able to make that easier for enterprises, from the agent's SDK or cloud developer platform. [00:38:12] These are scalable building blocks for all of you to be able to start thinking about embedding AI more easily. [00:38:17] We're also thinking about how to be able to have more context-specific evals that go into every single model evolution. [00:38:25] Really exciting, cool story. [00:38:27] One of our customers actually created their own coding language just in the last couple of months. [00:38:32] Of course, Claude, Sonnet 3.7, didn't have awareness of that. [00:38:36] And we built that into our LRL environment and allowed Claude to really be pre-trained that way before Sonnet 4.5. [00:38:44] And that, I think, now allows us to make much more intelligent models. [00:38:48] But no matter how intelligent the model is, if we don't have all the data architecture that's available, it's much more difficult for businesses to be able to build these powerful agents. [00:39:01] And so, while building the agent is becoming easier now, it's the data foundation that is a critical piece of the conversation. [00:39:09] Right. [00:39:10] And then, when you are working with these customers, how are, I guess, your most innovative customers, like the ones that are either getting this right or maybe they're on the cusp of getting this right? [00:39:19] Like, what are they doing, essentially, to unlock that data? [00:39:22] And how does real-time data play a role? [00:39:24] Yeah, absolutely. [00:39:25] I think our most sophisticated customers are recognizing, like I was saying, that, first of all, the foundation is really key. [00:39:32] But there's also a lot of relationships and patterns and nuance within data that we often forget. [00:39:37] And that's kind of what I was referencing when I was saying that there's this need for both structured and unstructured data. [00:39:42] Structured data, I think, is pretty clear. [00:39:43] We all know where that is, but I think when you're looking at how much AI can really empower each of your teams, you need to focus on how are your teams actually working. [00:39:53] That could be Slack teams, whichever, you know, team you decide to go for. [00:39:57] It can be for email, for calendaring. [00:39:59] It can be a lot of the ways in the systems that your teams are interacting. [00:40:03] That unstructured data actually has a lot of information that is extremely relevant to Claude or any frontier model. [00:40:10] And that's really where we need to recognize that that's real-time data tied to all the historical context and data that we also need. [00:40:17] And that feeds into a more context-aware agent. [00:40:21] Tying that together is really extremely powerful. [00:40:24] I already mentioned Agents SDK, which is a more universal, easier way to build generic agents across a number of different lines of business, not just coding, [00:40:34] which I think is one that we've jumped so quickly to, and that really allows us to harness the power of Claude, right? [00:40:42] Because, again, models are becoming wildly more intelligent, but they're also able to perform tasks on a much longer horizon. [00:40:50] We were just testing. [00:40:51] We have this game that's like Claude plays software engineer. [00:40:53] It's a very exciting internal game. [00:40:55] Everyone knows that Claude's played Pokemon. [00:40:56] A lot of fun things that we play around with at Anthropic, and Claude can actually do so for 30 hours. [00:41:04] I'm pretty sure Claude time of 30 hours is not even Rachel's time of 30 hours, probably like six months of my time. [00:41:11] And being able to tie that together, especially in the greater ecosystem of MCP, where it has access to that real-time data, allows us to, again, build that context that allows the agent to be able to reason and execute much more efficiently in a multi-layer agentic architecture. [00:41:31] Yeah. [00:41:32] And you mentioned MCP. [00:41:33] I'm sure that's a huge topic, a conversation. [00:41:34] It's a huge topic. [00:41:36] Definitely a big conversation at Confluent as well, something we're very excited about. [00:41:39] And as Jay mentioned, we just announced our fully managed MCP server today for a real-time context engine. [00:41:46] I guess, like, how do you see real-time and sort of MCP fitting together with the broader data and AI ecosystem? [00:41:53] Yeah. [00:41:53] First of all, congratulations. [00:41:54] That's really, really exciting. [00:41:56] I think we're excited to be able to see that announcement ourselves at Anthropic. [00:42:00] I think we love to call it MCP. [00:42:02] Tech loves abbreviations. [00:42:04] I think when we were slowing down and just talking about the model context protocol, something that Anthropic released about a year, [00:42:09] and a half ago, that is now globally leveraged, which is really, I think, exciting for us, because for us, we really care about how models retrieve and reason a bunch of different sources of external data and how that streams into that particular model. [00:42:22] And also, I think, something that we forget is that MCP allows AI to be able to stay extremely accurate, not just from a real-time standpoint, but from that historical level that we were talking about before, [00:42:33] and ensures that it's grounded in really relevant information. [00:42:36] I've been talking a lot about how AI has become incredibly more intelligent, and there's a lot of power in that. [00:42:43] But how do you make sure that it is contextually relevant? [00:42:46] And that's, I think, where we're seeing that shift to much more operational use cases being born. [00:42:53] And I think MCP is absolutely the key there. [00:42:56] And I think together, you know, especially this release from Confluent, it allows all of this live context streams directly into Claude, or I suppose other frontier models, to be able to power the agents itself. [00:43:08] And that, I think, is a really magnificent partnership. [00:43:13] Yeah, absolutely. [00:43:14] So, I guess, you know, we're running out of time, but before we wrap, what is your sort of one prediction for enterprise AI going into 2026? [00:43:21] Yes. [00:43:21] I would love to predict Transformers. [00:43:23] I have mentioned that Bumblebee is my favorite, but unfortunately, in San Francisco, we just have Waymo, though I won't be, I don't know what's going to happen on that side of the fence. [00:43:33] But I would say when I'm thinking specifically about enterprise AI, I don't think that we're going to be having dialogues like this anymore. [00:43:42] I genuinely think that we're going to get to the place where a lot of enterprises, a lot of businesses are going to understand that context is king. [00:43:50] And so that data architecture layer, that foundation is going to be where we're going to build all of our houses, aka agents. [00:43:56] And so the question really more now is how can you start building today? [00:44:01] Because I do think, if not, a lot of us will be playing catch up later, and that's how I'm encouraging quite a bit of our customers. [00:44:08] And so what I would really encourage everyone to think about is look at all of the different teams that you work with, or your own team in particular. [00:44:17] I definitely am encouraging my own team to think about it this way as well. [00:44:21] And think about all of the very amorphous problems that you have that have a lot of variables in it and have a lot of different data streams feeding into that. [00:44:30] That creates quite a bit of just like unknown and think about like how you can better connect all of that into real time data tied to any of the historical patterns that you will also need and think about building an agent there. [00:44:46] I think about how to be able to have a better agentic workflow, as I was saying, is never easier to build an agent. [00:44:52] And that's where I think that in a year, we'll really be able to move away from this conversation of accuracy. [00:44:58] Like, are we actually operationally accurate when it comes to building with AI and actually into like meaningful expansion? [00:45:05] That's where I think the power unlock is, is really being able to expand each of us in a much more meaningful way. [00:45:11] Well, awesome. [00:45:12] Thank you so much for being here. [00:45:13] We're really excited about our partnership with Anthropic and all the things that we're going to do together. [00:45:17] Next, I want you all to hear from our customers, Marriott and Metronome, on their data streaming AI journey. [00:45:24] Thank you so much. [00:45:25] Thank you. [00:45:25] Please welcome VP of Application Development and Architecture at Marriott International. [00:45:54] Rajesh Khandasamy. [00:46:07] Good morning, everyone. [00:46:09] I'm very excited to be here with you all. [00:46:13] I'm Rajesh Khandasamy, Vice President at Marriott. [00:46:17] I manage entire data streaming and also lead customer domain, [00:46:23] which includes identity, access management, and profile management, [00:46:29] serving over 248 million members across the globe. [00:46:35] Today, I will share how data modernization behind Marriott Bonvoy [00:46:41] elevates the guest experience from shop to book to stay. [00:46:50] We at Marriott have been in the business since 1927. [00:46:56] We started our journey as a root beer stand in Washington, D.C. [00:47:00] From that small beginning, we grew to 9,600 properties, 30-plus brands in 144 countries. [00:47:13] Our vision is to become world favorite travel company and not just being a hospitality leader. [00:47:20] You probably know Marriott Bonvoy, our loyalty program. [00:47:25] If any of you are not member yet, please let us know. [00:47:28] We will be very happy to sign you up for it. [00:47:31] Delivering your customer loyalty is a high-stake business for Marriott. [00:47:38] Because 70% of the revenue comes from the most loyal and valuable customers like you. [00:47:48] Our customers have high expectation on us to know who you are, [00:47:54] what is your need and optimal ways to engage you. [00:48:00] So, having a customer data real-time is key to our success. [00:48:05] And data streaming become a critical part of our business. [00:48:14] Imagine, you are a traveler who loves to plan your trip last minute, like me. [00:48:22] When you make that reservation within 24 hours of your arrival, [00:48:28] we get your reservation, understand you and your preferences. [00:48:34] You may like that extra towel or that special wine for your occasion. [00:48:40] We get all the relevant data from multiple systems, [00:48:44] enrich the data, [00:48:46] and make the data as mature data product, [00:48:50] and deliver that to the property in real time. [00:48:55] Our associates in property [00:48:57] prepare well in advance [00:48:59] to welcome you [00:49:01] and to provide the best experiences [00:49:04] you deserve during your stay. [00:49:07] In the past, [00:49:09] this used to be a fragmented experience. [00:49:13] Because our batch process took from 24 to 48 hours [00:49:19] to deliver the most valuable data. [00:49:25] Let me take a step back. [00:49:28] You are familiar with this. [00:49:30] We had this spaghetti, complex, point-to-point integration architecture. [00:49:36] Even a simple schema change was a big deal [00:49:40] because our reservation system integrates [00:49:44] to 20 to 30 different integration points. [00:49:49] And data moving from point A to B to C [00:49:52] introduced the lack of auditability [00:49:55] and data redundancy all over, [00:49:57] which you are all familiar with. [00:50:00] So, we embarked our data streaming journey [00:50:03] by introducing our first ever Kafka platform. [00:50:09] And we quickly moved into Confluent platform [00:50:11] and to the Confluent Cloud, [00:50:13] which powers our business today. [00:50:19] we have this simple, unflexible architecture built for a scale. [00:50:27] By moving to Confluent Cloud, [00:50:30] our engineers are no more spending their valuable time [00:50:35] managing a cluster or upgrading a platform. [00:50:40] Instead, they love developing products, tools, [00:50:46] real-time analytics using Flink [00:50:48] and advancing our governance [00:50:51] by securing our PIA data using CSFLE. [00:50:56] That enabled our platform as a self-services capability [00:51:02] to developer and business communities. [00:51:07] Because of that, [00:51:09] we have 900 critical business transactions [00:51:12] and 2,000 plus consumers [00:51:16] are consuming those data in real-time, [00:51:18] which could not be possible in the past. [00:51:24] At Marriott, [00:51:27] our core products are what you see in the property. [00:51:30] Examples, [00:51:31] room, spa, and golf. [00:51:35] And emerging products comes from the partnership. [00:51:38] Our co-brand cards, [00:51:40] airline partnership, [00:51:41] Uber, Starbucks, [00:51:42] and we have 100 plus partnership across the globe. [00:51:48] Our Bonvoy members are not just earning a points. [00:51:52] They earn points [00:51:53] while they're taking a ride on Uber [00:51:56] or having a sip of coffee in the Starbucks. [00:52:02] Previously, [00:52:04] any partner integration used to take [00:52:06] six to nine months. [00:52:09] Too slow for our growth [00:52:11] by introducing data streaming [00:52:14] and using connector ecosystems [00:52:17] and enabling all of our API [00:52:21] as a self-services capability, [00:52:24] we brought that time down to six weeks. [00:52:28] Yes, you heard. [00:52:29] Six weeks. [00:52:31] Amazing [00:52:32] that this technology can help [00:52:36] when business and technology moves together. [00:52:40] We are in the tail end [00:52:42] of the biggest digital [00:52:44] and technology transformation [00:52:46] in Marriott history. [00:52:49] From the beginning, [00:52:51] we had one simple architectural principles. [00:52:55] No surprise. [00:52:57] Every data has to be real-time. [00:53:00] And I am so excited [00:53:02] what that transformation [00:53:04] can bring to our guests, [00:53:07] to our associates, [00:53:09] and to the property owner [00:53:10] in the next chapter. [00:53:15] Before I close, [00:53:17] let me leave with [00:53:18] three key takeaways. [00:53:21] Number one. [00:53:23] Make your business-critical transaction [00:53:25] real-time [00:53:26] and operationalize them. [00:53:30] Number two. [00:53:32] Once you have that real-time data, [00:53:35] enrich it. [00:53:36] Make the data [00:53:37] as more matured data product [00:53:39] using technology like Flink. [00:53:44] Number three. [00:53:46] Apply AI on top of that [00:53:48] to provide [00:53:50] hyper-personalized experience [00:53:52] to the most valuable customer. [00:53:56] And I'm not done it. [00:53:58] This one is very close to my heart. [00:54:01] Develop ecosystems [00:54:02] and tools [00:54:03] for your developer [00:54:06] and business communities. [00:54:08] When you empower them, [00:54:11] they are going to make you look good today [00:54:13] and in the future. [00:54:19] Marriott data streaming journey [00:54:22] has become [00:54:23] a core part of who we are. [00:54:25] And I would like to take this moment [00:54:27] to thank Confluent [00:54:29] for their amazing partnership [00:54:31] throughout this journey. [00:54:33] And I would like to thank [00:54:34] all of you [00:54:35] for driving this community forward. [00:54:39] Together, [00:54:39] we are not just streaming a data. [00:54:43] We are streaming a trusted scale. [00:54:45] Thank you so much [00:54:47] for listening to us today. [00:54:48] Please welcome [00:55:04] Chief Technology Officer [00:55:06] at Metronome, [00:55:08] Cosmo Wolf. [00:55:09] Hey, all. [00:55:19] I'm Cosmo, [00:55:20] the CTO of Metronome. [00:55:21] And I'm super excited to be here [00:55:23] to talk about [00:55:23] the Confluent Metronome journey [00:55:25] for two reasons. [00:55:27] The first reason [00:55:28] is Confluent [00:55:29] is a big customer [00:55:31] of Metronomes, [00:55:32] which means [00:55:33] if you've ever [00:55:33] gotten an invoice [00:55:35] from Confluent, [00:55:36] which I imagine [00:55:36] most of you have, [00:55:38] or you've logged [00:55:39] into your Confluent account [00:55:40] to check your usage [00:55:41] or your spend, [00:55:43] or you've taken advantage [00:55:44] of the Confluent Cloud [00:55:46] free trial, [00:55:47] those are all experiences [00:55:48] that have been powered [00:55:49] in some ways [00:55:50] by the Metronome's [00:55:51] monetization infrastructure. [00:55:54] But the second reason [00:55:54] I'm excited to be here, [00:55:56] which is even more important, [00:55:57] is it's not an exaggeration [00:55:58] to say that [00:55:59] Metronome would not exist [00:56:00] without streaming data [00:56:03] and thus without Kafka [00:56:04] and without Confluent. [00:56:07] And I want to explain [00:56:08] why that is. [00:56:10] But first, [00:56:11] let's talk a teeny bit [00:56:12] about what Metronome does. [00:56:14] Metronome builds [00:56:14] monetization infrastructure, [00:56:16] which is to say, [00:56:18] we build the tools [00:56:19] and the APIs [00:56:19] and the infrastructure [00:56:21] that allow you [00:56:22] to collect value [00:56:23] and monetize your product. [00:56:25] And we do that [00:56:26] for some of the fastest moving [00:56:27] and most innovative [00:56:28] software companies in existence, [00:56:29] including many companies [00:56:30] which you use [00:56:31] or know day-to-day, [00:56:33] including Confluent, [00:56:34] like I already mentioned, [00:56:36] OpenAI, [00:56:37] Anthropic, [00:56:38] Databricks, [00:56:39] DataStax, [00:56:40] and many more. [00:56:42] Metronome is powering [00:56:43] the core monetization [00:56:44] for all of these companies. [00:56:47] And monetization [00:56:47] has been changing [00:56:48] a lot recently. [00:56:50] 10 or 15 years ago, [00:56:51] it was not uncommon [00:56:52] to be able to build [00:56:53] a billion-dollar company [00:56:54] off of a simple [00:56:56] subscription, [00:56:57] or a seat [00:56:57] or like simple [00:56:59] license monetization model. [00:57:02] But that is changing. [00:57:04] That world is over. [00:57:05] Today, [00:57:06] how you monetize [00:57:07] is a core part [00:57:08] of your product. [00:57:09] And almost every company [00:57:11] is experimenting [00:57:11] with consumption-based billing, [00:57:13] hybrid models, [00:57:14] credits, [00:57:15] prepaid commits, [00:57:16] and other [00:57:16] more complex [00:57:18] monetization strategies. [00:57:20] And even in the most [00:57:21] competitive markets, [00:57:22] we're seeing more [00:57:22] novel approaches, [00:57:23] like outcome-based pricing [00:57:25] or dynamic pricing. [00:57:27] And all of these [00:57:29] are real-time data challenges. [00:57:32] Because if your billing [00:57:36] is now flexible [00:57:37] and can no longer [00:57:38] sit downstream [00:57:39] of your product, [00:57:40] if user spend can change [00:57:42] minute-to-minute, [00:57:43] hour-by-hour, [00:57:44] day-by-day, [00:57:46] users will demand [00:57:47] more visibility [00:57:48] into your spend. [00:57:49] So they will seek out [00:57:50] products with better [00:57:51] spend visibility [00:57:52] and controls [00:57:53] that they have access to. [00:57:55] And they need those [00:57:55] controls to be [00:57:57] real-time. [00:57:59] And on top of spend [00:58:00] controls, [00:58:01] if you're rolling out [00:58:02] flexible pricing, [00:58:04] you might need [00:58:04] entitlements, [00:58:05] like Confluence [00:58:06] free trial, [00:58:07] I already mentioned, [00:58:07] or OpenAI's [00:58:09] free credits [00:58:10] when you sign up. [00:58:11] And those [00:58:12] entitlements [00:58:12] require timely gating, [00:58:14] so you can cut users off [00:58:15] once they've used [00:58:16] their allocated amount [00:58:18] of whatever product [00:58:19] you are selling [00:58:19] with the entitlement. [00:58:21] And if you're selling [00:58:23] some sort of [00:58:23] consumption-based product, [00:58:25] you need to protect [00:58:26] yourself from fraudulent [00:58:27] usage, [00:58:27] especially in low-margin [00:58:29] businesses like AI [00:58:30] or Infrasass. [00:58:31] So you want to make sure [00:58:32] you can cut users off [00:58:34] when they've used [00:58:35] more of the product [00:58:35] than they have paid for [00:58:36] and wait for them [00:58:37] to re-up [00:58:38] so you're not losing [00:58:38] money on fraudulent usage. [00:58:40] And on top of all of that, [00:58:42] the rest of the business, [00:58:43] like go-to-market [00:58:44] and finance, [00:58:45] needs frequent, [00:58:46] up-to-date visibility [00:58:47] into spend [00:58:48] so that they can do [00:58:49] their jobs, [00:58:50] upselling, [00:58:51] closing the books, [00:58:52] intervening and [00:58:53] course-correcting [00:58:54] customers who are [00:58:54] using more or less [00:58:55] than you expected, [00:58:56] et cetera. [00:58:58] And this means [00:58:59] every company now [00:59:02] has to become [00:59:03] a real-time data company. [00:59:06] And this is basically [00:59:08] what Metronome has built. [00:59:10] We're a real-time [00:59:11] observability system, [00:59:13] but for your revenue data. [00:59:15] So our customers [00:59:15] are streaming in events [00:59:17] as they occur [00:59:18] in their systems, [00:59:19] things like users signed up [00:59:21] or API inference occurred [00:59:23] or broker minute occurred [00:59:25] or bytes sent [00:59:27] or received or stored. [00:59:28] They're streaming that data [00:59:29] into Metronome [00:59:30] in real-time. [00:59:31] And then we are computing [00:59:32] their revenue [00:59:34] and financial data [00:59:34] based on complex pricing [00:59:36] and packaging. [00:59:37] Think like, [00:59:38] what is the business model? [00:59:39] How much do the products cost? [00:59:40] What is the product catalog? [00:59:41] And then we are streaming them back [00:59:43] real-time visibility, [00:59:45] control, [00:59:45] and insights [00:59:46] into their revenue data. [00:59:49] And this is a difficult problem. [00:59:51] In the classic kind of [00:59:52] streaming iron triangle, [00:59:55] normally you could give up [00:59:56] one of correctness, [00:59:57] availability, [00:59:58] or latency. [00:59:59] Maybe you decide, [01:00:00] you know, [01:00:01] I want to be a highly available, [01:00:03] very low latency product, [01:00:04] but I don't need [01:00:05] to be exactly correct. [01:00:06] I could do at least once [01:00:07] or at most once delivery. [01:00:08] Or maybe you want [01:00:09] very correct, [01:00:11] highly available processes, [01:00:12] but you can give up [01:00:13] a little bit of latency [01:00:14] and run a slower process [01:00:15] to error correct or whatnot. [01:00:17] But at Metronome, [01:00:18] we can't sacrifice [01:00:19] any of those three. [01:00:22] We have to run [01:00:23] a highly available, [01:00:25] incredibly low latency, [01:00:27] and exactly end-to-end correct [01:00:29] transactional observability system. [01:00:32] And we have to run that [01:00:33] at extreme scale. [01:00:36] We're currently doing [01:00:37] hundreds of billions of events, [01:00:39] all with sub-second latency, [01:00:42] feeding into compute [01:00:43] hundreds of millions [01:00:43] of invoices per day. [01:00:45] And again, [01:00:46] all of this end-to-end [01:00:48] has to be 100% accurate [01:00:50] because we are [01:00:50] a financial product. [01:00:54] And we're still a startup. [01:00:55] We have less than [01:00:56] 150 total employees now, [01:00:58] and we got started, [01:00:59] we only had three engineers. [01:01:01] But we still needed [01:01:02] to solve those problems [01:01:03] on day one. [01:01:04] Our product wouldn't have worked [01:01:05] without that core data [01:01:07] streaming architecture. [01:01:10] And that's what Confluent [01:01:11] enabled for us. [01:01:12] From the very beginning, [01:01:13] we were able to build [01:01:14] a production-grade, [01:01:15] real-time financial platform [01:01:16] with a truly tiny team. [01:01:18] And then over the past years, [01:01:20] we've been able to scale that [01:01:21] with a lean team [01:01:22] to be truly world-scale [01:01:24] infrastructure [01:01:24] that we operate today. [01:01:28] And looking ahead, [01:01:30] you are all going to need [01:01:32] to solve the same problems. [01:01:35] Less nimble companies [01:01:35] are always out-competed [01:01:37] by faster-moving companies. [01:01:40] And I would guess [01:01:40] every one of your companies [01:01:41] in the audience [01:01:41] is thinking about [01:01:42] how to incorporate AI [01:01:44] into your products. [01:01:45] Or if you're not, [01:01:46] your competitors are, [01:01:48] and you will feel [01:01:49] the competitive pressure [01:01:49] to evolve your pricing [01:01:50] and packaging [01:01:51] because you are [01:01:52] in the same market [01:01:53] with companies that are. [01:01:56] Which means [01:01:56] you're going to have [01:01:57] to become [01:01:58] a real-time data [01:01:59] and a real-time [01:01:59] monetization company [01:02:00] whether you like it or not. [01:02:04] And this is what [01:02:05] Confluent and Metronome [01:02:07] exists to help you with. [01:02:09] So you can focus [01:02:09] on building your product [01:02:10] and not on building [01:02:11] the core monetization [01:02:12] and data systems [01:02:14] in-house. [01:02:16] And me [01:02:17] and some people [01:02:18] from the Metronome team [01:02:18] are here all week [01:02:19] and if you have [01:02:20] any questions [01:02:21] about anything [01:02:22] real-time [01:02:23] whether it be [01:02:24] monetization or otherwise [01:02:25] please come find us. [01:02:26] We'd be happy to chat. [01:02:27] And thank you Confluent [01:02:28] for having me here [01:02:29] to talk about our story. [01:02:39] Please welcome [01:02:40] Chief Product Officer [01:02:42] of Confluent [01:02:43] Sean Klaus. [01:02:48] Well, a big thank you [01:02:56] to Anthropic, [01:02:57] Metronome [01:02:57] and Marriott. [01:02:58] What a tremendous [01:02:59] set of use cases. [01:03:01] You know, [01:03:01] I think that our partner [01:03:02] Anthropic [01:03:02] said it best [01:03:03] when they said [01:03:04] that without real-time data [01:03:05] there really is [01:03:06] no practical production AI. [01:03:09] And you saw that [01:03:10] validated through [01:03:10] some real-world proof. [01:03:12] You heard about [01:03:13] the data streaming platform [01:03:14] powering AI [01:03:15] for real-time billing [01:03:16] for digital natives [01:03:17] and for powering [01:03:19] really intimate [01:03:19] guest experiences [01:03:20] at the largest [01:03:22] hotel brand [01:03:22] in the world. [01:03:24] But we see [01:03:25] those same benefits [01:03:26] play out [01:03:26] at our almost [01:03:27] 6,000 customers [01:03:28] all over the world. [01:03:30] And we want [01:03:30] those benefits [01:03:31] for you [01:03:31] in whatever projects [01:03:32] you're currently [01:03:33] working on [01:03:33] in whatever industry [01:03:35] you are in. [01:03:37] Our mission for you [01:03:38] is actually [01:03:39] really simple. [01:03:40] We want to make [01:03:40] streaming ubiquitous, [01:03:42] get all of your data [01:03:43] moving [01:03:44] from the simplest [01:03:45] application [01:03:45] to the AI agents [01:03:47] that are transforming [01:03:48] your business. [01:03:49] And when streaming [01:03:50] is ubiquitous [01:03:51] we can find [01:03:52] a better way [01:03:53] to build. [01:03:54] We can shift [01:03:54] governing and processing [01:03:56] left, [01:03:57] closer towards the source, [01:03:59] do that work once [01:04:00] and then reuse the data [01:04:01] to power our AI agents, [01:04:03] our analytics [01:04:04] and our real-time [01:04:05] customer experiences. [01:04:08] Stream once, [01:04:09] use everywhere. [01:04:11] But today [01:04:11] instead of telling you how [01:04:12] we'd like to show you [01:04:14] how through a series [01:04:15] of different demos. [01:04:17] We're going to take a look [01:04:17] at a really traditional, [01:04:19] interesting streaming [01:04:20] use case, [01:04:21] ride-sharing. [01:04:23] But given that we're here [01:04:24] in New Orleans, [01:04:25] the big easy, [01:04:26] as they say, [01:04:27] let's make this [01:04:27] a little bit fun. [01:04:35] Yep, we're doing it. [01:04:37] We're doing it. [01:04:38] So, [01:04:40] we're... [01:04:40] Thank you. [01:04:42] Thank you. [01:04:42] We're going to be looking [01:04:46] at the data streaming platform [01:04:48] at River RoboTaxi. [01:04:50] There are startups [01:04:51] specializing in, [01:04:52] you guessed it, [01:04:53] autonomous riverboat taxis [01:04:55] on the Mississippi. [01:04:56] Now, [01:04:56] they're getting really popular [01:04:57] here in the South, [01:04:59] and they're looking [01:04:59] to use AI [01:05:00] to optimize [01:05:01] vessel routing, [01:05:03] staffing, [01:05:04] and logistics [01:05:04] around surges in demand, [01:05:07] often around large events. [01:05:09] Think things like [01:05:10] when the Saints [01:05:11] are playing in town, [01:05:12] Mardi Gras, [01:05:13] or even events like this one, [01:05:14] current here [01:05:15] at the convention center. [01:05:17] Now, [01:05:17] like I said, [01:05:18] this is a really interesting [01:05:19] streaming use case [01:05:20] because we have to receive [01:05:21] a lot of real-time information. [01:05:23] We have to recognize [01:05:24] that there is a surge in demand. [01:05:26] We have to analyze [01:05:26] what is causing that, [01:05:28] and then respond [01:05:28] by rerouting vessels [01:05:30] to meet that demand. [01:05:31] Now, [01:05:32] we're going to show you [01:05:33] how the data streaming platform [01:05:34] lets River [01:05:36] unleash streaming [01:05:37] for every use case [01:05:39] cost-effectively [01:05:40] all the way [01:05:41] across their business, [01:05:42] and then [01:05:43] they can power [01:05:43] their AI agents, [01:05:45] applications, [01:05:46] and experiences [01:05:47] with governed, [01:05:49] processed, [01:05:50] trustworthy, [01:05:52] real-time data. [01:05:53] And by the time [01:05:53] we're done, [01:05:54] the data mess [01:05:55] that plagues [01:05:55] organizations like River, [01:05:56] and maybe some of you, [01:05:58] will be a thing [01:05:59] of the past. [01:06:01] Now, [01:06:01] let's start [01:06:01] with that first step, [01:06:03] streaming. [01:06:04] Now, [01:06:04] given we are here [01:06:05] at Current, [01:06:06] I'm not going to need [01:06:06] to explain to all of you [01:06:07] that streaming [01:06:08] is the living, [01:06:10] beating heart [01:06:11] of the data streaming platform. [01:06:13] And that's all [01:06:13] thanks to Apache Kafka. [01:06:15] The energy [01:06:16] in this community [01:06:17] has never been higher. [01:06:19] You, [01:06:20] this community, [01:06:21] is the reason [01:06:21] that Kafka has gone [01:06:23] from being a niche [01:06:24] data technology [01:06:25] to being a foundational [01:06:26] part of the way [01:06:27] modern application [01:06:29] and indeed businesses [01:06:30] are built. [01:06:33] Today, [01:06:34] over 150,000 organizations [01:06:36] use Kafka, [01:06:38] and there are more [01:06:38] than 1,800 Kafka meetups [01:06:41] all over the world. [01:06:42] And Kafka itself [01:06:43] keeps getting better [01:06:44] and better. [01:06:45] Earlier this year, [01:06:46] the Kafka community [01:06:47] celebrated a really [01:06:48] major milestone [01:06:49] with the release [01:06:50] of Apache Kafka 4.0. [01:06:53] And yet, [01:06:54] time has marched on, [01:06:55] and we've also already [01:06:56] seen the release [01:06:57] of Apache Kafka 4.1. [01:06:58] These have been [01:06:59] some of the biggest releases [01:07:00] the Kafka community [01:07:02] has ever seen. [01:07:04] We've completed [01:07:04] the migration [01:07:05] from Zookeeper [01:07:06] to Kraft. [01:07:07] In fact, [01:07:08] soon, [01:07:08] there will be [01:07:08] no supported version [01:07:10] of Kafka [01:07:11] that works [01:07:12] with Zookeeper. [01:07:13] We've also seen [01:07:14] dramatic improvements [01:07:15] to consumer rebalancing [01:07:16] to make it simpler [01:07:18] and faster [01:07:19] so that your applications [01:07:20] are just better [01:07:21] and more reliable. [01:07:24] But the Kafka community [01:07:25] has been cooking up [01:07:25] something really big. [01:07:28] And so I'm excited [01:07:28] to share [01:07:29] that queues for Kafka [01:07:30] is now available [01:07:31] in open preview [01:07:33] in Apache Kafka. [01:07:39] This is a really big deal. [01:07:41] With queues, [01:07:42] developers can access [01:07:43] message queue semantics [01:07:45] directly on top of Kafka [01:07:46] without the need [01:07:47] for separate messaging [01:07:48] or queuing systems. [01:07:50] You don't need [01:07:51] to do anything. [01:07:53] Every Kafka topic [01:07:54] is available [01:07:55] as normal, [01:07:56] as a stream, [01:07:57] or as a queue, [01:07:59] or both at once. [01:08:01] So you can now power [01:08:02] really classical use cases, [01:08:03] things like asynchronous APIs, [01:08:05] job scheduling, [01:08:07] order management, [01:08:08] notifications, [01:08:09] and many more [01:08:10] on top of the reliability [01:08:11] and stability [01:08:13] of the Kafka platform. [01:08:14] Now, in many organizations, [01:08:16] Kafka has already replaced [01:08:17] a lot of their need [01:08:18] for message queues, [01:08:19] so this makes completing [01:08:20] that transition [01:08:21] to one reliable foundation [01:08:23] really trivial. [01:08:24] Please scan the QR code [01:08:25] if you'd like to learn [01:08:26] more about that. [01:08:28] But what's next? [01:08:29] Well, you heard me say, [01:08:30] we want to make [01:08:31] streaming ubiquitous. [01:08:32] With streaming, [01:08:33] we have found [01:08:33] a better way [01:08:34] to build more reliable [01:08:35] customer experiences. [01:08:38] But sadly, [01:08:38] some organizations [01:08:39] still use streaming [01:08:41] sparingly. [01:08:43] Why is that? [01:08:45] Well, it's because [01:08:45] they fear [01:08:46] that it will be too costly [01:08:47] or risky [01:08:48] to expand streaming [01:08:50] to all of their data [01:08:51] and all of their data [01:08:51] use cases. [01:08:52] And there's a few reasons [01:08:53] for that, [01:08:54] but none of them [01:08:54] are laws of physics. [01:08:57] First, organizations [01:08:58] often over-provision [01:08:59] their streaming infrastructure [01:09:00] to make certain [01:09:01] that they can meet [01:09:02] peak demand. [01:09:04] But that means [01:09:04] that in general, [01:09:05] 50% of the infrastructure [01:09:07] cost is wasted [01:09:08] underutilized technology. [01:09:10] Storage can be [01:09:11] a real pain. [01:09:12] Teams are forced [01:09:13] to choose between [01:09:14] short retention [01:09:15] on local disks [01:09:16] or trying to figure out [01:09:17] how to bolt on [01:09:18] external object storage [01:09:19] to retain data [01:09:20] for longer in Kafka. [01:09:22] Networking can be [01:09:23] unbelievably painful. [01:09:25] In some scaled [01:09:26] cloud implementations, [01:09:28] networking ends up [01:09:28] representing over 80% [01:09:30] of the total cost [01:09:31] of ownership of Kafka. [01:09:33] And at the end [01:09:34] of it all, [01:09:35] self-managed Kafka [01:09:35] really can't provide [01:09:37] the SLA governance [01:09:38] and security capabilities [01:09:39] that enterprises [01:09:40] really need. [01:09:43] You could just try [01:09:44] using a managed Kafka. [01:09:45] Maybe you could try [01:09:46] the one from your [01:09:46] cloud service provider. [01:09:48] But at the end [01:09:49] of the day, [01:09:50] all that really changes [01:09:51] is where your invoice [01:09:51] comes from. [01:09:53] What looks like [01:09:54] for one-click [01:09:54] managed Kafka [01:09:55] is really just [01:09:56] Apache Kafka [01:09:56] on the back end [01:09:57] and all of the scaling, [01:09:59] tuning, [01:10:00] and firefighting [01:10:01] ends up being left [01:10:02] to you and your team. [01:10:05] Now, that's why [01:10:05] we created Quora. [01:10:07] It's our Apache Kafka [01:10:08] engine built [01:10:09] for the cloud. [01:10:10] And Quora lets us [01:10:11] offer a cloud-native [01:10:13] Kafka experience [01:10:14] to our almost 6,000 [01:10:15] customers all over the globe. [01:10:18] With Quora, [01:10:19] we offer [01:10:19] a serverless, [01:10:21] elastic, [01:10:22] instantly auto-scaling [01:10:23] Kafka [01:10:24] that delivers [01:10:25] 10 times better [01:10:26] tail latencies, [01:10:27] a four nines SLA, [01:10:29] and infinite storage [01:10:30] out-of-the-box. [01:10:34] But I don't have [01:10:35] a lot of time [01:10:36] on stage this morning, [01:10:37] so if you'd like [01:10:38] to learn more [01:10:38] about the architectural [01:10:39] advantages of Quora, [01:10:41] I'd encourage you [01:10:41] to check out [01:10:42] the QR code [01:10:42] on the screen. [01:10:44] Instead, today, [01:10:44] I really want to focus [01:10:45] on one key aspect [01:10:46] of Quora [01:10:47] that's critical [01:10:48] to our mission [01:10:48] of making streaming [01:10:50] ubiquitous, [01:10:51] and that's cost. [01:10:53] Our mission [01:10:54] is to make streaming [01:10:55] cost-effective [01:10:56] for every use case [01:10:57] everywhere, [01:10:58] no ifs, [01:10:59] ands, [01:11:00] or buts. [01:11:01] Whether you need [01:11:02] a massive throughput [01:11:03] topic with [01:11:04] tremendously low latency, [01:11:06] or you just need [01:11:07] the least expensive way [01:11:08] to stream data [01:11:09] from A to B, [01:11:10] we're focused [01:11:10] on delivering [01:11:11] the exact right [01:11:12] streaming option [01:11:12] for you [01:11:13] and your cost profile. [01:11:15] And that's what [01:11:16] Quora was built [01:11:16] to deliver. [01:11:17] With Quora, [01:11:18] we have removed [01:11:19] costs at every [01:11:20] single layer [01:11:20] of the stack. [01:11:22] Quora instantly [01:11:23] auto-scales [01:11:24] so that 50% [01:11:25] of over-provisioned [01:11:26] infrastructure [01:11:26] is a thing [01:11:27] of the past. [01:11:29] Quora ships [01:11:30] with diskless Kafka [01:11:31] out of the box. [01:11:32] Infinite storage [01:11:33] is just there. [01:11:35] And if you choose [01:11:36] to use object storage [01:11:37] as the replication layer, [01:11:38] all of that cross-availability [01:11:39] zone networking traffic [01:11:41] that can end up being [01:11:42] so much of the networking [01:11:44] cost of Kafka [01:11:44] is a thing [01:11:45] of the past. [01:11:47] Speaking of networking, [01:11:49] with our brand new [01:11:49] private network interface [01:11:51] powered by AWS's [01:11:52] Elastic Network [01:11:53] Interface technology, [01:11:55] you can get secure, [01:11:57] private cloud networking [01:11:58] for 50% less cost. [01:12:01] And all of that [01:12:02] comes with [01:12:02] the 4.9's SLA [01:12:03] and security [01:12:04] and governance capabilities [01:12:05] that enterprises [01:12:06] really need. [01:12:08] But how exactly [01:12:09] does this show up [01:12:10] for all of you? [01:12:11] Simple. [01:12:12] Options. [01:12:15] We have a variety [01:12:16] of different, [01:12:18] sorry, [01:12:19] we have a whole fleet [01:12:19] of different fully [01:12:20] auto-scaling clusters [01:12:21] that meet exactly [01:12:23] the performance, [01:12:23] capabilities, [01:12:24] and cost profile [01:12:25] of your data streaming needs. [01:12:27] But what you really need [01:12:28] to hear is [01:12:29] no matter what [01:12:30] your use case is, [01:12:31] we have [01:12:32] the highest performance, [01:12:34] most capable, [01:12:35] lowest cost streaming option, [01:12:37] no ifs, [01:12:38] ands, [01:12:38] or buts. [01:12:38] And we operate [01:12:39] all of these cluster types [01:12:41] at massive scale, [01:12:42] so it's normally [01:12:43] the hardest part, [01:12:44] managing Kafka [01:12:45] becomes the easiest part. [01:12:48] And this isn't theoretical. [01:12:50] We operate [01:12:51] over 30,000 clusters [01:12:53] across over 100 cloud regions, [01:12:57] sustaining more than [01:12:58] 8.2 trillion messages a day [01:13:01] with that 4.9's SLA [01:13:02] and rock-solid resilience [01:13:04] that I was talking about [01:13:05] a moment ago. [01:13:06] And that's how, [01:13:07] at our customers' [01:13:08] products from Indeed [01:13:09] to SAS and beyond, [01:13:11] we reduce their streaming cost [01:13:13] by between 40% and 70%, [01:13:14] but more importantly, [01:13:16] we unlock streaming [01:13:17] for all of their data use cases. [01:13:19] And we can do the same [01:13:21] for you, too. [01:13:22] Whatever your data [01:13:23] or workload is, [01:13:24] we will get it streaming [01:13:25] at scale, [01:13:26] and we will make [01:13:26] the economics undeniable. [01:13:29] But, like I said, [01:13:30] let's see this in action. [01:13:32] River have their vessel information [01:13:34] stored at an Oracle DB. [01:13:35] They have vessel telemetry [01:13:37] and ride requests [01:13:38] coming in from some [01:13:39] custom-built applications. [01:13:40] So to walk us through [01:13:41] that demo, [01:13:42] please welcome back [01:13:43] Sean Falconer. [01:13:54] All right. [01:13:55] Well, thank you, Sean. [01:13:56] It's good to be back up here. [01:13:58] It looks like it's going to be [01:13:59] the Sean and Sean show [01:14:00] where one of us [01:14:01] spells our name correctly [01:14:02] and the other one [01:14:03] thinks they do. [01:14:04] I'll leave it up to you [01:14:04] to decide who it is, [01:14:05] but it's this guy. [01:14:07] Well, I think we're going [01:14:08] to have to agree to disagree [01:14:08] on that one, [01:14:09] but at least we're both [01:14:10] well-dressed. [01:14:11] Which is the most important part. [01:14:12] All right. [01:14:13] Well, I'm excited [01:14:14] to be back up here [01:14:14] to take all of you [01:14:16] through some demos. [01:14:16] Hopefully, you're excited [01:14:17] as well. [01:14:19] Let's get started. [01:14:20] Let's cue the demo. [01:14:21] So, Sean talked about [01:14:23] how we want to keep [01:14:24] data streaming [01:14:25] cost-effective for everyone. [01:14:27] And we're going to put [01:14:28] that to the test [01:14:29] with our river rowboat [01:14:30] taxi company. [01:14:31] So, here we have [01:14:32] two very different [01:14:33] types of data. [01:14:34] We have our ride requests, [01:14:35] which are time-sensitive. [01:14:36] We don't want people [01:14:37] waiting around [01:14:38] for their boat. [01:14:39] So, we're going to use [01:14:40] enterprise clusters for this. [01:14:41] They're built for speed. [01:14:43] And then on the flip side, [01:14:43] we have things like [01:14:44] telemetry data [01:14:45] that's continuously streaming [01:14:46] from our vessels. [01:14:47] Here we have a huge [01:14:48] volume of data. [01:14:49] We don't necessarily [01:14:50] need blazing speed, [01:14:51] but we want to be able [01:14:52] to do this [01:14:52] and store that data [01:14:53] without essentially [01:14:54] breaking the bank. [01:14:55] So, we're going to use [01:14:55] freight clusters for that, [01:14:56] which I'm setting up now. [01:14:57] They're built for this. [01:14:59] Both are elastic [01:15:00] and autoscale, [01:15:01] so you only pay [01:15:01] for what you need. [01:15:02] And they're both [01:15:03] being protected [01:15:04] using private networking [01:15:05] through our new [01:15:06] private networking interface. [01:15:08] This is based [01:15:08] on the same AWS [01:15:09] core primitives [01:15:10] that protect their services, [01:15:12] deployed within your VPC [01:15:13] to minimize [01:15:14] data transfer costs. [01:15:16] All right. [01:15:16] So, we have our [01:15:17] cluster set up. [01:15:18] Now, we actually [01:15:19] need some data. [01:15:20] And we're going to start [01:15:20] with our vessel catalog data, [01:15:22] which represents [01:15:23] the ship information [01:15:24] that we have. [01:15:25] This is currently stored [01:15:26] in Oracle. [01:15:27] So, in order to get [01:15:28] the data from Oracle [01:15:29] into Confluent Cloud, [01:15:30] we're going to use [01:15:31] our new Oracle Extreme [01:15:32] CDC connector, [01:15:34] which I'm setting up here. [01:15:35] Pretty simple. [01:15:36] Basically, point it [01:15:36] at your database, [01:15:37] put in your credentials. [01:15:39] Once I have that, [01:15:40] I select continue. [01:15:42] And then, [01:15:42] the next step [01:15:43] is just to choose [01:15:44] the Kafka topic [01:15:45] that I want to stream [01:15:45] this data into. [01:15:47] And there we go. [01:15:48] This is fully managed. [01:15:49] Just click through, [01:15:51] launch the connector, [01:15:51] and we have a data [01:15:53] flowing from Oracle [01:15:54] directly into Confluent Cloud. [01:15:56] And we can take a look [01:15:56] at what this looks like. [01:15:58] So, we'll pull up [01:15:58] one of the sample data [01:15:59] and this essentially [01:16:00] represents our ships [01:16:01] within our fleet. [01:16:03] All right. [01:16:03] So, that's our database data. [01:16:05] Now, for the fun stuff, [01:16:06] our real-time data [01:16:07] coming from [01:16:08] all of our fleet. [01:16:09] So, every one of our ships [01:16:10] is a moving data generator. [01:16:12] We're talking fuel levels, [01:16:14] temperature of the engine, [01:16:16] speed, [01:16:16] GPS location, [01:16:17] all coming from these ships [01:16:18] in extremely high volume. [01:16:21] So, [01:16:22] when you have, [01:16:23] essentially, [01:16:24] thousands of different sensors [01:16:26] sending millions [01:16:27] of different data, [01:16:27] and we'll pull up [01:16:28] an example of one of these [01:16:29] in a second, [01:16:30] the big question, [01:16:31] of course, [01:16:32] is how do you prevent [01:16:33] bad data [01:16:34] from potentially poisoning [01:16:36] the entire system? [01:16:37] How do you make sure [01:16:38] that you have data quality [01:16:39] and force [01:16:39] from the very beginning? [01:16:40] So, there's an example [01:16:41] of some of our [01:16:42] telemetry data. [01:16:43] The way that we solve [01:16:44] this problem [01:16:44] is through schemas. [01:16:46] These act as a contract [01:16:47] for your data [01:16:48] where you can essentially [01:16:50] enforce the rules [01:16:51] and the shape of your data [01:16:52] the way that you want [01:16:53] that data constructed. [01:16:54] So, this is an example [01:16:55] of our vessel telemetry. [01:16:56] So, that's our operational data. [01:16:58] The next step [01:16:59] is our ride request. [01:17:00] And for this, [01:17:01] we're going to go over [01:17:01] to the app. [01:17:02] So, when you're a rider [01:17:04] with River, [01:17:05] you essentially go into the app, [01:17:06] select the location, [01:17:09] request a boat. [01:17:11] There we go. [01:17:11] And then, [01:17:12] behind the scenes, [01:17:13] this request gets written [01:17:14] into a Kafka topic [01:17:16] called ride requests. [01:17:17] We'll pull up an example [01:17:18] here in a second. [01:17:19] And there we go. [01:17:21] In less than five minutes, [01:17:23] we're streaming. [01:17:24] We have our database connected. [01:17:25] We have our telemetry data [01:17:26] flowing in. [01:17:27] We also have our ride requests. [01:17:29] So, what used to be [01:17:30] one of the hardest things [01:17:32] to essentially manage [01:17:33] and configure Kafka [01:17:34] is now one of the easiest. [01:17:36] And with that, [01:17:36] I'm going to hand it back [01:17:37] to Sean. [01:17:40] All right. [01:17:41] Thank you. [01:17:42] Thank you, Sean. [01:17:43] So, you'll notice [01:17:45] how Sean was able [01:17:46] to mix and match [01:17:47] different cluster types [01:17:48] to meet the different needs [01:17:49] of his data, [01:17:50] but to do so cost effectively. [01:17:53] I hope you also saw [01:17:54] how Sean was easily able [01:17:55] to connect to his Oracle database. [01:17:57] And that was because [01:17:58] he was able to take advantage [01:17:59] of Connect [01:18:00] and Confluent Hub, [01:18:01] which is the richest collection [01:18:03] of connectivity [01:18:04] in the entire streaming ecosystem. [01:18:07] There are hundreds [01:18:07] of connectors [01:18:08] on Confluent Hub. [01:18:09] We're constantly adding more, [01:18:11] but also our partners [01:18:12] and community members [01:18:13] like you [01:18:13] are constantly adding [01:18:14] more connectivity [01:18:15] to the ecosystem. [01:18:17] All right. [01:18:18] I said earlier [01:18:18] that we're aiming for ubiquity. [01:18:20] And to truly be ubiquitous, [01:18:21] we need to reach data [01:18:22] everywhere that it lives. [01:18:24] And it turns out [01:18:25] that about 50% [01:18:26] of the world's data [01:18:27] lives on-premise [01:18:28] or in private clouds. [01:18:30] So, we need the right [01:18:31] streaming capabilities [01:18:32] for those deployment models, too. [01:18:34] Now, almost exactly [01:18:35] a year ago today, [01:18:37] WarpStream [01:18:37] joined the Confluent family. [01:18:40] And they have [01:18:41] a fantastic streaming offer [01:18:42] for customers [01:18:43] who prefer a [01:18:44] bring-your-own-cloud [01:18:45] deployment model [01:18:46] thanks to their purpose-built [01:18:47] bring-your-own-cloud [01:18:48] control plane. [01:18:50] And for customers [01:18:51] who really need to run [01:18:52] all of their streaming environment [01:18:54] inside their own data center [01:18:55] or to run it entirely [01:18:57] within their virtual [01:18:57] private cloud, [01:18:58] Confluent Platform [01:18:59] is perfect for that, too. [01:19:01] Well, these are a great start, [01:19:02] but we still think [01:19:03] we can do better. [01:19:05] Kafka's success [01:19:06] can create sprawl. [01:19:07] It starts with one team, [01:19:08] then it's tens of teams. [01:19:10] Before you know it, [01:19:11] you have a bunch [01:19:11] of different topics [01:19:12] and clusters spread [01:19:13] across different business units [01:19:14] or departments. [01:19:16] And the central platform team [01:19:17] is usually the one [01:19:18] who carries all that weight. [01:19:20] They're the one [01:19:20] who scale the fleet, [01:19:21] apply patches, [01:19:23] manage access, [01:19:23] keep everything compliant [01:19:25] and resilient. [01:19:26] They're understaffed, [01:19:27] underappreciated. [01:19:29] We know the pain well. [01:19:30] We have been running Kafka [01:19:32] for hundreds of customers [01:19:33] for over a decade. [01:19:35] So we want something better [01:19:36] for those teams. [01:19:38] We're introducing today [01:19:39] Confluent Private Cloud. [01:19:47] With Confluent Private Cloud, [01:19:50] you can self-manage [01:19:51] a cloud-like Kafka environment [01:19:53] within your virtual private cloud [01:19:55] or data center. [01:19:57] Think things like [01:19:57] multi-tenancy, [01:19:59] scaling, [01:20:00] performance, [01:20:01] patches, [01:20:02] upgrades, [01:20:02] and a lot more. [01:20:04] And Confluent Private Cloud [01:20:05] isn't just about management. [01:20:07] It also extends [01:20:08] all the way to governance too [01:20:09] because CPC ships with Gateway, [01:20:12] our smart, [01:20:13] policy-aware, [01:20:14] front door for Kafka. [01:20:16] With Gateway, [01:20:17] you can centralize authentication [01:20:19] and authorization. [01:20:21] You can enforce [01:20:22] schemas and quotas [01:20:24] and you can smartly [01:20:26] route traffic [01:20:27] by policy. [01:20:28] And what does that mean? [01:20:30] Well, that means [01:20:31] that you can prevent [01:20:32] one team's [01:20:32] bad access patterns [01:20:34] or client configuration [01:20:35] from impacting [01:20:36] another team. [01:20:37] You can securely [01:20:38] control [01:20:39] and enable [01:20:40] sharing of data [01:20:41] across teams [01:20:42] or even [01:20:43] with your external partners. [01:20:45] And client failover [01:20:47] and disaster recovery [01:20:48] is dramatically simplified. [01:20:51] Now Confluent Private Cloud [01:20:52] is really perfect [01:20:53] for organizations [01:20:54] in highly regulated environments [01:20:56] or even managed service providers [01:20:58] where a platform team [01:21:00] is managing Kafka [01:21:00] on behalf of many other teams. [01:21:03] But not only [01:21:04] does Confluent Private Cloud [01:21:05] make it much easier [01:21:07] to manage Kafka at scale, [01:21:08] it actually makes it [01:21:09] less expensive too. [01:21:11] In fact, [01:21:12] Confluent Private Cloud [01:21:13] can reduce your cost [01:21:14] of operations [01:21:15] and infrastructure [01:21:16] for Kafka [01:21:16] by over 50%. [01:21:18] How's that? [01:21:20] Well, first, [01:21:21] in Confluent Private Cloud [01:21:22] we're shipping [01:21:23] the intelligent replication layout [01:21:25] from Quora [01:21:26] in the cloud [01:21:27] into your on-prem environment. [01:21:29] And that means [01:21:29] that your brokers [01:21:30] are up to 10 times faster. [01:21:34] Second, [01:21:35] we also ship CPC [01:21:37] with intelligent bin packing, [01:21:39] which means you get [01:21:40] better resource scheduling, [01:21:41] maximize the utilization [01:21:42] of your underlying infrastructure [01:21:44] and reduce the need [01:21:45] for this over-provisioning [01:21:46] that plagues [01:21:47] most self-managed Kafka environments. [01:21:49] And finally, [01:21:50] with Gateway, [01:21:51] you get the ability [01:21:52] to enforce standards [01:21:54] at one central point [01:21:55] and reduce the need [01:21:57] for cluster sprawl [01:21:58] to reduce blast radius [01:21:59] of individual teams' usage [01:22:01] across your Kafka environments. [01:22:04] And it's going to get better. [01:22:05] Soon, [01:22:06] we're going to ship [01:22:06] the entire logical multi-tenant layer [01:22:09] of Quora in the cloud [01:22:11] into the on-prem environment [01:22:12] so you can get [01:22:13] even better utilization [01:22:14] of your Quora environments as well. [01:22:16] Sorry, [01:22:16] your Confluent Platform Cloud environment as well. [01:22:18] Now, [01:22:18] Confluent Private Cloud [01:22:20] is GA today. [01:22:23] Check out the QR code [01:22:24] to learn more about that. [01:22:31] But honestly, [01:22:32] we recognize that [01:22:33] most large organizations [01:22:33] don't have [01:22:34] just one deployment model. [01:22:36] and that's what has made [01:22:37] our cluster linking technology [01:22:39] so successful. [01:22:41] With cluster linking, [01:22:42] you can stream data [01:22:43] between your on-prem, [01:22:45] private cloud environments [01:22:47] and the public clouds, [01:22:48] AWS, [01:22:49] GCP, [01:22:49] and Azure. [01:22:50] You can stream the data [01:22:51] seamlessly, [01:22:52] within milliseconds, [01:22:54] offset preserving [01:22:55] with zero knobs [01:22:56] or dials to configure. [01:22:57] That's what we mean [01:22:58] by ubiquitous streaming. [01:23:01] But we also want [01:23:01] to make it easier [01:23:02] for you to understand [01:23:03] this complicated data flow [01:23:04] and infrastructure. [01:23:06] So today, [01:23:07] we're excited to announce [01:23:08] our brand new [01:23:09] Unified Stream Manager. [01:23:15] With Unified Stream Manager, [01:23:17] you can manage [01:23:18] all of your confluent clusters [01:23:19] from one place. [01:23:21] You can see [01:23:21] all of your schemas, [01:23:23] all of your metrics, [01:23:24] all of your lineage [01:23:24] across all of your streams [01:23:26] from a single pane of glass. [01:23:29] Unified Stream Manager [01:23:30] is shipping GA today [01:23:32] connecting to [01:23:33] Confluent Platform 8.1. [01:23:34] You can scan the QR code [01:23:36] to learn more. [01:23:37] All right. [01:23:38] Let's get back [01:23:38] to our use case. [01:23:40] We've got our data connected [01:23:41] and our streams [01:23:42] are up and running. [01:23:43] But if we're really [01:23:44] going to power AI agents, [01:23:46] we don't just need [01:23:47] real-time data. [01:23:48] We need data [01:23:48] that is trustworthy [01:23:49] and has been contextualized [01:23:51] to the business problem [01:23:53] we're trying to solve [01:23:53] or the decision [01:23:54] we want the AI to make. [01:23:57] And processing data [01:23:58] with Flink [01:23:58] is the key [01:23:59] to making rich real-time data [01:24:01] an actual reality. [01:24:03] With Flink, [01:24:04] you can work with, [01:24:05] shape, [01:24:06] and transform your data [01:24:07] in real-time [01:24:09] without first needing [01:24:10] to land it [01:24:11] in a database [01:24:11] or a data lake. [01:24:13] And in Confluent Cloud, [01:24:14] our serverless Flink offering [01:24:16] lets you instantly auto-scale [01:24:18] a Flink job [01:24:18] to any scale [01:24:19] to any scale [01:24:19] that you need, [01:24:20] making it trivial [01:24:21] for you to filter, [01:24:23] join, [01:24:24] aggregate, [01:24:25] enrich, [01:24:26] or standardized data [01:24:27] while it is still [01:24:28] in motion, [01:24:29] on the move. [01:24:31] In the last year [01:24:32] since we GA'd [01:24:32] our Flink offerings, [01:24:33] they've proven themselves [01:24:34] to be the most successful products [01:24:36] we have ever launched [01:24:37] as a company. [01:24:39] But how does that all [01:24:40] play out at River? [01:24:41] Well, obviously, [01:24:42] they have a bunch [01:24:42] of information [01:24:43] they need instantly [01:24:44] to run their business, [01:24:45] vessel location, [01:24:47] weather, [01:24:47] local events. [01:24:49] But real-time signals [01:24:50] aren't by themselves enough. [01:24:52] They also need information [01:24:53] about historical ridership patterns [01:24:55] or seasonal demand patterns. [01:24:57] That information [01:24:57] is not generated [01:24:58] in real-time, [01:24:59] but it's needed [01:25:00] in real-time [01:25:01] at the point of decision. [01:25:03] Now, luckily, [01:25:04] the data streaming platform [01:25:05] already has [01:25:06] all of the data [01:25:06] that we need [01:25:07] to power these use cases. [01:25:08] And that's where [01:25:09] Confluent Intelligence [01:25:10] comes in. [01:25:11] As Jay shared a moment ago, [01:25:13] with Confluent Intelligence, [01:25:14] you can unlock [01:25:15] your fresh, [01:25:16] governed data [01:25:17] for instant use [01:25:18] inside your AI agents. [01:25:20] So let's see it in action. [01:25:22] Sean, [01:25:22] show us some AI agents. [01:25:24] All right. [01:25:25] Sounds great, Sean. [01:25:26] So to recap, [01:25:28] we want to use AI [01:25:29] to coordinate logistics. [01:25:31] Things like [01:25:31] our staffing levels, [01:25:33] how many vessels [01:25:34] that we have available [01:25:35] based on the number [01:25:36] of ride requests [01:25:36] that we're seeing [01:25:37] at any given moment. [01:25:38] And we're going to break [01:25:39] that up into smaller chunks [01:25:40] just so that you can see [01:25:42] how achievable [01:25:43] an AI use case [01:25:44] like this truly is. [01:25:45] To start, [01:25:46] we're going to apply [01:25:47] anomaly detection [01:25:48] for our ride requests [01:25:50] in real time. [01:25:52] Let's cue the demo. [01:25:54] All right. [01:25:54] So we're going to use [01:25:55] built-in anomaly detection [01:25:56] available in Confluent Intelligence. [01:25:58] This is the easiest place [01:25:59] to start with applying AI [01:26:01] to your real-time data. [01:26:02] What we're doing here [01:26:03] is we're using [01:26:04] anomaly detection [01:26:04] over essentially [01:26:05] periods of time [01:26:06] as we analyze [01:26:07] this real-time stream [01:26:08] of ride requests. [01:26:09] and we're looking [01:26:10] for deviations [01:26:10] in that pattern [01:26:11] that represent [01:26:12] these types of anomalies. [01:26:14] Once we recognize that, [01:26:15] we can flag it [01:26:16] and this allows us [01:26:17] to act proactively, [01:26:18] potentially dispatching [01:26:19] new boats to the area [01:26:20] rather than being reactive [01:26:22] and dealing with this problem [01:26:24] down the stream. [01:26:26] Well, once we've run the, [01:26:28] we have that job running, [01:26:29] we can take a look [01:26:30] at our anomalies detected [01:26:31] and we see that [01:26:32] there's actually been [01:26:33] an anomaly. [01:26:33] There is a surge [01:26:35] in the number of ride requests [01:26:36] that are happening [01:26:36] right now. [01:26:38] Of course, [01:26:38] anomaly detection [01:26:39] is just the start. [01:26:42] There's a lot more work to do. [01:26:43] So, let's go back to slides [01:26:45] to talk about what's next. [01:26:48] All right, [01:26:49] so we're going to take [01:26:49] this a step further. [01:26:50] We want to be able [01:26:51] to take our anomalies [01:26:52] and cross-check them [01:26:54] with local events [01:26:55] that are happening [01:26:55] in the area. [01:26:56] You know, [01:26:56] as Sean said, [01:26:57] you know, [01:26:58] perhaps you have something [01:26:58] like Mardi Gras happening [01:27:00] or maybe there's a band [01:27:01] going up Canal Street [01:27:02] that's blocking access [01:27:03] to your hotel room [01:27:04] like there was [01:27:04] at my hotel last night. [01:27:06] Or it could be that, [01:27:07] you know, [01:27:08] everybody is on a rush [01:27:09] to get to Current [01:27:10] to be at the keynote [01:27:11] right now [01:27:11] and they all want [01:27:12] to take riverboats [01:27:13] because that's the fastest [01:27:14] way to get around. [01:27:15] Whatever it is, [01:27:16] essentially, [01:27:17] we want to know [01:27:18] not just that [01:27:19] something is happening [01:27:20] but why it's happening [01:27:21] and how long [01:27:22] it might be taking [01:27:23] so that we can [01:27:24] intelligently respond [01:27:25] and mitigate the issue [01:27:26] using AI agents. [01:27:28] In order to do this, [01:27:29] we're going to use [01:27:29] our streaming agents capability. [01:27:32] Let's queue the demo. [01:27:34] All right, [01:27:35] so usually when people [01:27:36] think of an AI agent, [01:27:37] they think about [01:27:37] a chatbot waiting around [01:27:38] for a user to ask a question. [01:27:40] And what we're talking [01:27:41] about here [01:27:41] is something [01:27:42] completely different. [01:27:43] This is an event-driven agent. [01:27:45] It's not prompted, [01:27:47] it's not reacting [01:27:48] to user prompts. [01:27:49] It's triggered, [01:27:49] essentially, [01:27:50] by the data itself. [01:27:51] And the way [01:27:52] that we make that possible [01:27:53] is we've built agents [01:27:54] as a core component [01:27:55] within Flink. [01:27:56] I just specify, [01:27:57] essentially, [01:27:57] the model that I'm using, [01:27:59] this example quad, [01:28:00] the prompt, [01:28:01] the tools that I need, [01:28:02] and then I run [01:28:03] that agent definition [01:28:04] directly in the Flink runtime. [01:28:06] In this case, [01:28:07] pulling in additional data [01:28:08] about the local events [01:28:10] that we're pulling, [01:28:11] essentially, [01:28:11] from a vector search, [01:28:12] as well as the vessels [01:28:14] that are nearby. [01:28:16] And once we run this, [01:28:17] we use the powerful reasoning [01:28:19] and model [01:28:19] to make decisions. [01:28:21] In this case, [01:28:22] if there's enough boats [01:28:23] in the area [01:28:23] to deal with the surge, [01:28:24] nothing happens. [01:28:25] But if not, [01:28:27] then we use [01:28:28] tool-calling capabilities [01:28:29] through MCP [01:28:30] to dispatch new boats [01:28:31] to the area. [01:28:32] What you end up with, [01:28:34] essentially, [01:28:34] is real-time, [01:28:36] proactive, [01:28:37] informed reasoning [01:28:38] about every decision. [01:28:41] And if we just pause [01:28:42] for a moment [01:28:42] and think about [01:28:42] what we just did, [01:28:44] is we built [01:28:44] a fully automated [01:28:45] fleet management system [01:28:46] using a combination [01:28:48] of anomaly detection [01:28:49] and an AI agent. [01:28:51] It's always on, [01:28:52] it's proactive, [01:28:52] it's streaming native. [01:28:54] All right, [01:28:54] let's go back to the slides [01:28:55] and talk about what's next. [01:28:57] All right, [01:28:57] so that was a great example, [01:28:59] but if you remember, [01:29:00] we want to use [01:29:01] the same real-time, [01:29:03] contextual data [01:29:04] to power AI [01:29:05] wherever you need it, [01:29:07] not just within Confluent. [01:29:08] So imagine that a rider [01:29:10] has requested a boat, [01:29:12] and despite our best efforts, [01:29:13] there's been a delay. [01:29:15] Now the rider goes [01:29:16] to the mobile app [01:29:17] to the customer service chatbot [01:29:19] and starts asking questions, [01:29:20] and that chatbot [01:29:21] is powered by an AI agent [01:29:22] that's built [01:29:22] outside of Confluent. [01:29:24] Well, in order for that [01:29:25] customer service chatbot [01:29:26] to be able to [01:29:27] intelligently respond [01:29:28] to the user, [01:29:29] it needs to understand [01:29:30] what is happening right now [01:29:32] in order to make [01:29:33] those types of decisions. [01:29:34] So in order to solve [01:29:35] this problem, [01:29:35] we're going to use [01:29:36] our real-time context engine. [01:29:37] Let's cue the demo. [01:29:40] All right, [01:29:40] so for this customer service [01:29:42] chatbot, [01:29:43] using an AI agent [01:29:44] that's been built [01:29:45] outside of Confluent, [01:29:46] in order to make [01:29:46] intelligent decisions, [01:29:47] it can't be doing that [01:29:48] based on stale data. [01:29:49] We need to connect it [01:29:50] to this real-time, [01:29:51] high-value context. [01:29:52] So in order to do that, [01:29:53] we're going to take [01:29:53] our Kafka topics, [01:29:55] some of the data [01:29:56] that we have flowing [01:29:56] through Kafka now [01:29:57] and enable them [01:29:58] for the context engine. [01:29:59] This turns those topics [01:30:01] into tools that can consume [01:30:04] through a fully-managed [01:30:05] MCP server. [01:30:06] And when I go into the app [01:30:07] and I ask about [01:30:08] the status of my ride, [01:30:10] behind the scenes [01:30:11] on the left, [01:30:11] you can see Claude's [01:30:13] chain-of-thought reasoning [01:30:14] as it interacts [01:30:15] with our MCP server, [01:30:16] first asking about [01:30:17] the data that's available [01:30:17] and then finding out [01:30:18] there's been a surge [01:30:20] and what action [01:30:21] was taken for it. [01:30:22] And then it can use that [01:30:23] to contextualize the response, [01:30:25] letting the user know [01:30:26] that there's been a delay [01:30:27] because thousands of people [01:30:28] are on their way [01:30:29] to current keynote right now [01:30:30] and we've dispatched [01:30:32] new boats to the area. [01:30:33] Please be patient. [01:30:36] So if we take a step back [01:30:38] and look at what we just built [01:30:39] over the last couple of minutes [01:30:40] is we powered [01:30:42] two completely different [01:30:43] AI use cases. [01:30:44] One, proactive [01:30:46] and system-triggered [01:30:47] using a combination [01:30:48] of anomaly detection [01:30:49] and an AI agent [01:30:50] and the other one, [01:30:52] reactive and user-driven [01:30:53] using the real-time [01:30:54] context engine. [01:30:55] All right, [01:30:56] let's go back to slides. [01:30:59] All right, [01:30:59] so we just saw [01:31:00] all three core tenets [01:31:02] of Confluent Intelligence [01:31:03] in action. [01:31:03] We started pretty simple [01:31:04] with applying anomaly detection [01:31:07] to the number of real-time [01:31:09] ride requests [01:31:11] that we were seeing [01:31:11] using Flink's built-in [01:31:13] machine learning capabilities. [01:31:15] And then from there, [01:31:16] we added a streaming agent [01:31:17] which allowed us [01:31:19] to essentially go deeper, [01:31:21] cross-reference that anomaly [01:31:23] with other information [01:31:24] about local events [01:31:25] and actually make [01:31:26] intelligent decisions. [01:31:26] And this was easy to build. [01:31:28] Of course, [01:31:28] the agent can access [01:31:29] all your real-time data [01:31:30] that's flowing within Confluent, [01:31:32] but it also has access [01:31:33] to data that exists [01:31:34] outside of Confluent [01:31:35] through external tables [01:31:36] as well as through MCP. [01:31:39] And when our AI system [01:31:40] was built outside of Confluent, [01:31:42] we used the real-time [01:31:44] context engine [01:31:44] to give instant access [01:31:45] access to your streams [01:31:47] so that to power [01:31:49] any AI system anywhere. [01:31:51] And what that all really means [01:31:52] is the same real-time, [01:31:55] contextualized, [01:31:56] trustworthy data [01:31:57] powers every AI system [01:31:59] you build, [01:31:59] whether it's a streaming agent [01:32:00] on Flink [01:32:01] or an external app [01:32:02] or agent. [01:32:04] If you want to learn [01:32:05] more about this, [01:32:06] you can scan this QR code, [01:32:08] and with that, [01:32:09] I'll hand it back to Sean. [01:32:09] All right. [01:32:17] Honestly, [01:32:17] how cool is that? [01:32:18] Streaming agents [01:32:19] in 15 lines of SQL. [01:32:21] Amazing. [01:32:22] So we're really excited [01:32:23] about our vision [01:32:23] for Confluent Intelligence, [01:32:25] and we're not the only ones. [01:32:27] Obviously, [01:32:27] you heard from Anthropic earlier, [01:32:28] but we'd like to take a moment [01:32:29] to highlight our other partners [01:32:31] along this journey. [01:32:33] So where are we at? [01:32:34] Back to the overall demo. [01:32:35] Well, now we've got our AI agents [01:32:37] up and making real-time decisions, [01:32:39] but obviously, [01:32:40] if you're a river, [01:32:40] you need to understand [01:32:41] the quality of those decisions [01:32:42] and what is going on, [01:32:44] something you typically do [01:32:45] in analytics, [01:32:46] so probably inside your data lake [01:32:48] or data warehouse. [01:32:50] Now, sadly, [01:32:50] traditionally, [01:32:51] it's actually been kind of hard [01:32:52] to get data from streaming [01:32:53] into your data lake [01:32:55] or data warehouse, [01:32:55] just like it used to be hard [01:32:57] to power your AI [01:32:58] with streaming data. [01:32:59] But as Jay shared earlier, [01:33:01] thanks to the emergence [01:33:02] of Delta and Iceberg, [01:33:04] these open table formats, [01:33:05] it doesn't have to be [01:33:06] that way anymore. [01:33:07] As Jay shared, [01:33:08] with Tableflow, [01:33:09] we've extended Kafka's storage engine [01:33:11] so that you can access [01:33:13] your Kafka topics [01:33:14] also as Iceberg or Delta tables. [01:33:17] So the exact same real-time, [01:33:20] reliable, reusable data [01:33:21] that is literally already powering [01:33:23] your AI agent [01:33:24] and your customer experiences [01:33:26] can now also immediately appear [01:33:28] in your data lake, [01:33:29] your data warehouse, [01:33:30] and all of your analytical tools. [01:33:32] There is no need for ETL, [01:33:34] ELT, [01:33:35] endless data prep. [01:33:37] It's just seamless data flow [01:33:38] from the streaming ecosystem [01:33:39] into your analytical ecosystem. [01:33:42] Honestly, [01:33:42] this is one of the most powerful technologies [01:33:44] I've ever had the opportunity [01:33:45] to release. [01:33:47] Now, Tableflow went GA [01:33:49] a little bit earlier this year [01:33:51] with support for AWS [01:33:53] and Iceberg, [01:33:54] but we have been hard at work [01:33:55] since then. [01:33:57] And actually, [01:33:57] today we have a cavalcade [01:33:58] of new announcements for you [01:34:00] around Tableflow. [01:34:01] Firstly, [01:34:02] we're announcing the GA [01:34:03] of support for Delta format [01:34:05] with Unity Catalog support. [01:34:07] We're also announcing the GA [01:34:09] of support for up-certs, [01:34:12] dead-letter queues, [01:34:13] and for encryption [01:34:14] across both Delta and Iceberg. [01:34:18] And finally, [01:34:18] we're super excited [01:34:19] to announce the early access release [01:34:21] of support of Tableflow [01:34:22] on Azure as well. [01:34:29] So, [01:34:30] what does all that mean? [01:34:31] Honestly, [01:34:32] from my perspective, [01:34:33] Tableflow once again [01:34:34] demonstrates the value [01:34:35] of shifting left, [01:34:37] moving governance [01:34:38] and processing [01:34:38] closer towards the source. [01:34:41] By doing that, [01:34:42] the data arrives clean [01:34:43] and well-governed. [01:34:45] That means [01:34:45] that it's schematized, [01:34:47] data that doesn't quite fit [01:34:48] can be routed [01:34:49] to a dead-letter queue, [01:34:50] and encryption exists [01:34:51] throughout. [01:34:53] And then, [01:34:54] what you're looking at [01:34:55] is not just a raw dump [01:34:56] of the Kafka changelog. [01:34:58] To be clear, [01:34:59] we take care of inserts, [01:35:01] upserts, [01:35:02] sorry, inserts, [01:35:03] updates, [01:35:04] compaction, [01:35:05] and table maintenance [01:35:06] so that you get [01:35:07] the keyed, [01:35:08] merged, [01:35:09] silver analytical tables, [01:35:10] which are the ones [01:35:11] that you really need. [01:35:12] So, for example, [01:35:13] if I was taking in [01:35:14] a stream of customer profile updates [01:35:16] with customer IDs [01:35:17] and individual fields [01:35:19] of a profile [01:35:19] that are being updated, [01:35:21] Tableflow can present [01:35:22] a table keyed [01:35:23] by customer ID [01:35:24] with complete customer profiles [01:35:26] easily. [01:35:28] and the best bit is [01:35:30] if your business logic changes, [01:35:32] you don't need to rewire [01:35:33] a bunch of pipelines, [01:35:34] you can just replay the data [01:35:35] from Kafka [01:35:35] and the tables will rebuild themselves. [01:35:39] Now, Tableflow [01:35:40] obviously gives you [01:35:42] better, faster data [01:35:43] in your analytical ecosystem, [01:35:45] but the best part is [01:35:46] it also reduces your costs too. [01:35:49] We see it time and again [01:35:50] at customers from Notion [01:35:52] to Siemens [01:35:52] and many more. [01:35:54] When you shift governance [01:35:55] and processing left, [01:35:57] you reduce your data quality issues [01:35:59] by up to 60%. [01:36:00] Your infrastructure costs [01:36:02] for analytical processing [01:36:03] are reduced by 30% [01:36:05] and more. [01:36:06] And your engineering teams [01:36:08] can move faster [01:36:08] than they ever have before. [01:36:11] All right. [01:36:11] Let's see Tableflow in action. [01:36:13] Sean, take it away. [01:36:14] All right. [01:36:15] Thanks, Sean. [01:36:16] So now River [01:36:17] wants to run some analytics [01:36:19] on how busy each zone is [01:36:21] and also the decisions [01:36:22] being made by their agents. [01:36:24] So they essentially [01:36:25] want to be able to [01:36:26] take the data [01:36:27] that exists in Compliant [01:36:28] and stream it into [01:36:30] Databricks [01:36:31] in order to run this analytics. [01:36:33] In order to solve that problem, [01:36:34] we're going to use Tableflow. [01:36:35] Let's queue the demo. [01:36:38] So Tableflow makes it [01:36:40] push-button simple [01:36:41] to turn your Kafka topics [01:36:43] into Delta Lake [01:36:44] or Iceberg tables. [01:36:46] And we want to take [01:36:46] the decisions being made [01:36:48] by our agents [01:36:49] and stream those [01:36:51] into Databricks. [01:36:52] So it's pretty simple. [01:36:53] We set up Delta Table [01:36:54] format there [01:36:55] and then I choose [01:36:55] my S3 bucket [01:36:56] where I want to store my data. [01:36:58] Click Continue. [01:36:59] Launch. [01:36:59] Launch Tableflow [01:37:00] and there with zero code, [01:37:02] zero pipelines to manage, [01:37:04] I'm moving data [01:37:05] from Coughlin Cloud [01:37:06] into Tableflow. [01:37:08] With Tableflow, [01:37:09] your Kafka topics [01:37:10] simply become your tables. [01:37:12] It's not just about the data. [01:37:13] It's also the schema [01:37:13] is kept in sync [01:37:15] within Unity Catalyst. [01:37:15] catalog. [01:37:16] And now we can go [01:37:17] over the Databricks [01:37:18] and start to explore [01:37:19] this information [01:37:20] and this data. [01:37:21] We can do that [01:37:21] using Genie AI [01:37:22] where I can express [01:37:23] my questions like [01:37:25] what are the number [01:37:26] of ride requests [01:37:27] we're seeing per zone [01:37:28] and it will take [01:37:29] that natural language [01:37:30] query [01:37:31] and convert it [01:37:32] into SQL [01:37:33] as you can see here [01:37:34] and actually execute that [01:37:35] and allow me [01:37:36] to explore this data [01:37:37] in a table [01:37:38] as well as [01:37:38] in some visual format. [01:37:40] and I can continue [01:37:41] to ask questions. [01:37:42] For example, [01:37:42] I can ask about [01:37:43] what actions [01:37:44] have been taken [01:37:45] by our agents [01:37:46] based on zones [01:37:47] so I can start [01:37:47] to assess the quality [01:37:48] of the AI decision making [01:37:50] and the AI investments [01:37:51] that I've made. [01:37:52] And the amazing thing [01:37:53] about when you [01:37:54] have high value, [01:37:57] clean data [01:37:58] continuously refreshed [01:38:00] through table flow [01:38:02] and near real time [01:38:03] within Databricks [01:38:03] is it unlocks [01:38:04] new use cases [01:38:05] for your business. [01:38:06] For River RoboTaxi [01:38:07] that means that [01:38:08] they can assess [01:38:08] their staffing needs, [01:38:10] their vessel capacity, [01:38:12] even determine [01:38:13] their ROI [01:38:13] for their AI investments [01:38:16] through streaming agents [01:38:17] and the anomaly detection [01:38:18] features they're using [01:38:18] within Confluent Intelligence. [01:38:20] And if you're using [01:38:21] things like Genie AI [01:38:22] you can now have [01:38:23] any user within your business [01:38:25] be able to explore that data. [01:38:26] No technical expertise [01:38:27] required. [01:38:29] All right. [01:38:30] And with that [01:38:30] let's go back to the slides. [01:38:33] All right. [01:38:33] So we're excited [01:38:34] to work with [01:38:35] all our partners [01:38:36] on this journey [01:38:37] to provide [01:38:39] and deliver [01:38:40] rich, reliable [01:38:41] real time data [01:38:42] to power any Lake House [01:38:44] or analytics platform. [01:38:45] Today [01:38:45] we're very excited [01:38:47] to announce [01:38:48] another major partnership [01:38:49] that opens up [01:38:50] an entirely new frontier [01:38:52] for your streaming data. [01:38:54] Confluent [01:38:55] and Salesforce [01:38:56] are partnering together. [01:38:58] Thank you. [01:39:02] These are two platforms [01:39:04] that sit at the heart [01:39:05] of how modern [01:39:07] enterprises operate. [01:39:09] We're connecting [01:39:09] the world's leading [01:39:10] data streaming platform [01:39:11] with the world's leading [01:39:12] CRM. [01:39:14] This is a pretty powerful idea. [01:39:16] I mean, think about [01:39:17] how many of your organizations [01:39:18] stream data through Kafka [01:39:20] and also use Salesforce [01:39:21] to power your customers, [01:39:23] your service, [01:39:24] and more. [01:39:25] Together, [01:39:26] we're unlocking [01:39:26] really a new generation [01:39:28] of AI-driven, [01:39:30] personalized [01:39:30] customer experiences. [01:39:33] So how do we make [01:39:34] all that happen? [01:39:36] Well, Tableflow powers [01:39:37] Salesforce Data Cloud [01:39:39] with trusted, [01:39:40] real-time data [01:39:41] across the enterprise. [01:39:42] This means that [01:39:43] every customer profile, [01:39:44] every dashboard [01:39:45] is continuously refreshed [01:39:47] and up-to-date. [01:39:49] In parallel, [01:39:50] Flink is analyzing [01:39:51] those real-time streams, [01:39:53] detecting patterns, [01:39:54] trends, anomalies, [01:39:56] key moments [01:39:56] in your business. [01:39:57] And when something [01:39:58] interesting happens, [01:40:00] it uses streaming agents [01:40:01] to invoke [01:40:02] and trigger [01:40:03] an agent-force workflow, [01:40:05] turning your AI investments [01:40:07] within Salesforce [01:40:08] into event-driven AI. [01:40:09] And together, [01:40:11] we make the Salesforce [01:40:12] vision of [01:40:13] agentic enterprise [01:40:14] possible. [01:40:15] To talk more about this, [01:40:17] please welcome [01:40:18] Gunther Hogleitner, [01:40:19] SVP of Engineering [01:40:20] at Salesforce. [01:40:33] Hey, Sean, [01:40:33] how's it going? [01:40:34] I have a small phone [01:40:35] to pick with you, actually. [01:40:36] Oh, no beads? [01:40:37] No beads, man. [01:40:38] You did not give me [01:40:38] the dress codes, [01:40:39] and now I feel [01:40:40] really underdressed [01:40:40] for the occasion. [01:40:41] All right. [01:40:42] Well, one of us [01:40:42] has to be a professional. [01:40:43] Well, thank you so much [01:40:44] for being here. [01:40:45] I really appreciate it. [01:40:46] We're very excited [01:40:47] about Salesforce. [01:40:48] So, you know, [01:40:49] Salesforce talks [01:40:50] a lot about [01:40:50] and is really leading [01:40:51] the conversation [01:40:52] around this idea [01:40:52] of the agentic enterprise [01:40:54] where you use, [01:40:55] you basically build [01:40:56] more autonomy, [01:40:57] power through AI agents [01:40:59] to automate more work [01:41:00] within your business. [01:41:02] If, from your experience, [01:41:03] like, why is real-time data [01:41:04] so central to essentially [01:41:06] unlocking that vision? [01:41:07] Yeah. [01:41:08] Yeah, so at Salesforce, [01:41:09] we've been able [01:41:09] to work with enterprises [01:41:11] across virtually every industry [01:41:13] to help them bring AI [01:41:15] into the flow [01:41:16] of how they serve [01:41:17] their own customers, [01:41:18] how they support [01:41:20] their employees, [01:41:22] or how they run [01:41:22] their operations. [01:41:24] With Agent Force, [01:41:25] our AI platform, [01:41:26] we've really been able [01:41:27] to touch every part [01:41:29] of the business. [01:41:31] The pace of innovation [01:41:32] that we've seen [01:41:33] has been absolutely amazing [01:41:34] and really stellar [01:41:35] to witness. [01:41:36] But as I suspect [01:41:38] many of you [01:41:39] here in the room, [01:41:40] we've also noticed [01:41:41] the pattern [01:41:41] that there is [01:41:43] a growing divide [01:41:44] between the type [01:41:45] of applications, [01:41:46] agentic applications, [01:41:47] that really deliver [01:41:50] elevated outcomes [01:41:51] and those [01:41:52] that kind of never [01:41:53] really fully achieve [01:41:54] the potential [01:41:54] that they set out [01:41:55] to achieve. [01:41:56] There's been [01:41:57] documented cases [01:41:58] in the industry [01:41:59] about this as well. [01:42:01] So digging into it, [01:42:02] Salesforce, [01:42:03] and as we heard here [01:42:04] from Jay and others today, [01:42:05] we think that context [01:42:06] is really the key [01:42:07] in bridging that divide [01:42:09] and getting over this. [01:42:10] And specifically, [01:42:11] real-time context, [01:42:13] real-time governed, [01:42:15] personalized context [01:42:16] makes all the difference there. [01:42:18] If you can provide [01:42:19] your agents [01:42:19] with the right level [01:42:20] of information [01:42:21] just at the time [01:42:22] that the interactions happen, [01:42:25] that's really [01:42:26] when you can achieve [01:42:27] these outcomes. [01:42:28] And so, yeah, [01:42:28] that's why we believe [01:42:29] that real-time context [01:42:31] really is the key here [01:42:33] and real-time [01:42:33] really plays [01:42:34] an outsized role [01:42:36] in that agentic enterprise. [01:42:38] Yeah, it's all about [01:42:38] the right data [01:42:39] at the right time. [01:42:40] And so, you know, [01:42:42] we announced this partnership, [01:42:45] Confluent, [01:42:46] Salesforce partnering together, [01:42:47] two of the most [01:42:48] widely used platforms [01:42:49] in their particular domain. [01:42:50] Confluent with data streaming, [01:42:52] Salesforce with [01:42:53] customer intelligence. [01:42:55] essentially with [01:42:56] Confluent being able [01:42:57] to bring Kafka data [01:43:00] into Salesforce [01:43:01] in high volume, [01:43:03] in real-time, [01:43:04] what are some [01:43:05] of the key unlocks [01:43:06] for your customers [01:43:06] in terms of value? [01:43:07] By the way, [01:43:08] I've been able to talk [01:43:09] to some of those [01:43:10] shared customers here [01:43:11] in the audience today already [01:43:13] and yesterday, [01:43:14] and I think they all echo [01:43:15] what we anticipated, [01:43:17] that putting the two [01:43:19] platforms together, [01:43:20] you really get more [01:43:21] than the sum of its parts. [01:43:22] You can really unlock [01:43:23] a lot of value. [01:43:24] I have a couple [01:43:25] of concrete examples [01:43:26] of, you know, [01:43:28] things that we've seen [01:43:29] where this kind of [01:43:29] combination of these platforms [01:43:31] makes a huge difference. [01:43:33] The first of them [01:43:34] is a customer support agent [01:43:37] in the telco industry [01:43:39] that we helped build. [01:43:41] There, [01:43:41] the key was really [01:43:43] to have the latest network, [01:43:46] load, [01:43:47] outage information [01:43:48] available to these [01:43:49] customer chatbots [01:43:52] right at the time [01:43:53] when the customer [01:43:54] is investigating [01:43:54] and trying to find out [01:43:56] what it is. [01:43:57] Minutes later, [01:43:58] hours later, [01:43:59] it doesn't help you [01:44:00] for this kind of application. [01:44:02] But having this context [01:44:03] in real-time, [01:44:04] they were able [01:44:05] to drive down [01:44:06] resolution times [01:44:08] by over a third, [01:44:08] which I think [01:44:09] is an amazing achievement. [01:44:11] Another example [01:44:12] that I think many of us [01:44:13] probably have run into [01:44:14] is personalized ads. [01:44:17] So we've all kind of seen [01:44:18] the type of ads [01:44:20] that you get served up [01:44:21] after you've already [01:44:22] made your purchase decision, [01:44:24] after you've already clicked buy, [01:44:26] or after you've moved [01:44:27] on to something else. [01:44:28] That is, you know, [01:44:29] not really useful. [01:44:30] That is annoying [01:44:31] in some cases. [01:44:32] And having that type [01:44:33] of real-time context [01:44:34] when these systems [01:44:36] are making those decisions [01:44:37] really can help you [01:44:38] overcome this [01:44:39] and really can, [01:44:40] as we've seen, [01:44:42] get you 60% [01:44:43] higher conversion rates, [01:44:44] drive up relevance, [01:44:45] and those types of things. [01:44:46] So I think we will see [01:44:47] more and more [01:44:47] of these systems, [01:44:48] and I think these systems [01:44:49] are really critical [01:44:50] to unlock that type of value. [01:44:51] Yeah, absolutely. [01:44:52] And, you know, [01:44:53] I joked earlier, [01:44:54] context is the word of the day, [01:44:56] maybe the word of the conference. [01:44:57] You can add it to your, [01:44:58] along with AI, [01:44:59] to your keyword bingo cards. [01:45:00] But, you know, [01:45:03] just so you kind of wrap up, [01:45:04] I wanted to ask, [01:45:05] from an engineering perspective, [01:45:07] and I know this is something [01:45:08] dear to your heart, [01:45:09] and I started my career [01:45:10] as an engineer, [01:45:10] now I mostly pretend to be one. [01:45:12] But what excites you [01:45:13] the most about [01:45:15] how these platforms, [01:45:16] you know, [01:45:17] work together? [01:45:18] Yeah, I started [01:45:18] as an engineer as well, [01:45:19] and the architecture [01:45:20] is actually very exciting to me. [01:45:22] I'm really pleased [01:45:23] with what we've settled on here. [01:45:25] So the architecture centers [01:45:27] around the iceberg [01:45:28] as an open and powerful format [01:45:30] for us to share information on. [01:45:32] The work you've done [01:45:33] with Tableflow [01:45:35] really, really helps [01:45:36] simplifying, [01:45:37] getting the data [01:45:38] from Kafka [01:45:38] into these iceberg tables. [01:45:41] And then the work [01:45:42] that we've done [01:45:42] at Salesforce [01:45:43] with zero copy [01:45:44] to get the data [01:45:46] without any additional [01:45:47] management overhead, [01:45:48] without any additional copies [01:45:49] or ingests [01:45:49] or any of that nature, [01:45:51] directly into Data360. [01:45:53] And then from there, [01:45:54] being able to activate it [01:45:55] across all of our channels, [01:45:58] be it Tableau, [01:45:59] be it Slack [01:46:01] or marketing, [01:46:02] sales, agents, [01:46:03] you can really bring [01:46:04] that data immediately [01:46:06] through this architecture [01:46:07] into those channels. [01:46:08] So that is super exciting. [01:46:10] If you pair that [01:46:11] with something like Flink [01:46:12] on top of it, [01:46:13] where you can just react [01:46:16] to business changes [01:46:17] very quickly [01:46:18] and then trigger [01:46:19] Agent4's workflows [01:46:21] directly from there, [01:46:23] I think that is super powerful. [01:46:25] And I think we'll see [01:46:26] this architecture [01:46:26] kind of replicated [01:46:27] many, many ways [01:46:28] and power a lot [01:46:29] of the applications [01:46:30] in future. [01:46:32] So I'm super excited [01:46:32] to see what our customers [01:46:33] will do with that partnership [01:46:35] and with that combined [01:46:36] architecture. [01:46:38] Absolutely. [01:46:38] And me too. [01:46:39] And well, thank you, [01:46:40] Guthrie, so much for coming. [01:46:41] We're really excited [01:46:42] about this partnership [01:46:43] and what we can do together [01:46:45] and the unlocks [01:46:45] for our customers. [01:46:46] And with that, [01:46:47] we're going to hand it back [01:46:48] to Sean to close this out. [01:46:50] Well, thank you so much. [01:46:51] Thank you. [01:46:58] And that's it, folks. [01:47:08] We solved the use cases [01:47:09] that we set out for River. [01:47:11] We made streaming ubiquitous. [01:47:13] We set every workload, [01:47:15] every data set in motion [01:47:16] at scale [01:47:17] and we did it cost effectively. [01:47:20] And using that streaming data, [01:47:21] we found a better way to build. [01:47:22] We shifted governance [01:47:23] and processing left. [01:47:25] That meant that we could do [01:47:26] the work once at the source [01:47:27] and reuse it [01:47:28] to power our AI agents, [01:47:30] our analytical use cases [01:47:31] and our real-time [01:47:33] customer experiences. [01:47:35] And what does that do? [01:47:36] Well, ultimately, [01:47:37] it converts real-time data [01:47:39] to real business advantage. [01:47:41] For River, [01:47:42] that means better vessel utilization, [01:47:44] better customer satisfaction [01:47:45] and more revenue coming in. [01:47:48] That is the power [01:47:49] of a data streaming platform [01:47:50] and we are really excited [01:47:51] to see what the data streaming [01:47:52] platform can do for you. [01:47:54] Thank you, everybody, [01:47:55] for joining us here [01:47:56] in New Orleans. [01:47:57] We hope you have [01:47:58] a fantastic day. Thank you.