РАСШИФРОВКА 🔗
speaker-0: Good morning, good afternoon, good evening. ⁓ I am Stephen Algrady. I'm the co founder of Red Monk, the developer focused ⁓ industry analyst firm. And we are here today to talk about, well, some some interesting intersections between ⁓ agents in the agentic AI era and data platforms where those two things come together. So to join me, ⁓ I am joined by Boris. Boris, would you introduce yourself briefly?
speaker-1: Sure. Hi hi everybody. My name is Boris Bierleg. I'm the Vice President of Industries at MongoDB and I'm the data side of the discussion today.
speaker-0: Excellent. And Roman, would you care to introduce yourself? I
speaker-2: ⁓ good morning, good evening, good afternoon, everyone. My name is Raman Jatkar. I'm head of product management at Intellect Design. ⁓ I lead the Purple Fabric ⁓ platform and I'm ⁓ you know, I'm going to be representing the AI agents and the platform play.
speaker-0: Awesome. All right. Sounds good. So on that note, ⁓ I won't spend a ton of time on this because a lot of this is probably understood from our from our audience's perspective. But obviously, you know, we we very much are in a a AI-driven agentic era. ⁓ and I won't get into some very long, convoluted, you know, sort of definition of of ⁓ agents, but essentially we have you know these autonomous agents backed by models that are operating ⁓ in some cases, you know, sort of from the from the get-go of of application generation in terms of code assists. you know, all the way out to performing, you know, sort of enterprise functions. And obviously the the sort of underlying substrate here and the backdrop to all of this is data. You know, and so what I wanted to talk about today is sort of the intersection of of you know sort of that era, if you will, that AI agentic era with the underlying data platforms that we have relied on for for years at this point. So to to start, ⁓ you know, Roman, you mentioned purple fabric. ⁓ can you tell us sort of what that is and and what it does? Can you give us sort of the the sort of intro to Purple Fabric?
speaker-2: Purple Fabric is an open business impact AI platform. It's a multi-agent AI platform. It has ⁓ you know particularly four key tea ⁓ tech stacks. The first one is the Enterprise Knowledge Garden. This is where this is the foundational intelligence where you're ingesting all your data across different data sources, ⁓ across different forms, whether it is structured, unstructured, and how do you translate all of that data into AI ready knowledge? Right. ⁓ of course there is an entire pipeline in terms of pre processing, applying different chunking strategies, re rank a model, and this is where MongoDB sits in as the key underpinning layer from a vector store perspective. Right. The second tech stack is the enterprise digital expert. This is where the agents live and breathe. You can build single agents, team of agents, automated workflows, ⁓ and basically build the multi agent solutions. ⁓ whether it is goal based, whether they are, you know, very well defined workflows and so forth, right? And we provide a very new code layer because we feel that ⁓ it, you know, this technology sits between business ⁓ and ⁓ you know, technology and a business technology should be technologists should be able to build these agents without really going through an engineering sprint. Right? The third tech stack which is very key, which is what we call as the L LM optimization hub. One important aspect is purple fabric is model agnostic. ⁓ We are ⁓ we we sort of have ⁓ models across different providers, the entropic models, ⁓ cloud perspective, the open AI models, Gemini models and so forth. And we provide benchmarking capabilities and all of this so you can actually benchmark your ⁓ you know response against the standard ground truth and measure your accuracy, latency and cost. And finally, of course, the governance layer, which ⁓ is by the way, embedded across all of these tech stacks, whether you look at it from an auditability standpoint, when you look at it from governance, guardrails, security, and so forth. So in effect, ⁓ in in in summary, ⁓ it purple fabric constitutes of these four tech stacks. And we the aim is ⁓ to help enterprises move from data to decisions. That's how I would summarize what Purple Fabric is.
speaker-0: Okay. Sounds great. And obviously sort of one of the foundations of any platform as we talked about at the top is the underlying data. And ⁓ so Boris, what can you tell us about where Mongo fits into Profabric?
speaker-1: Yeah, so it's really exciting. We are delivering really the data and the knowledge garden and what we learned out of over fifty client projects what we have done so far together, which is an amazing number when you think about everybody's doing some pilots and prototypes. Have the having fifty real clients together gives you a lot of experience. And what we learn is getting the knowledge into the knowledge garden, into the MongoDB level is absolutely critical. And at the end these data are more critical for the end solution when you think about ⁓ Raman talked about financial services. No, your client onboarding fraud. If you look at those, the data were more critical than the models. And this is an unbelievable brutal statement. When you look at people by now they these fights, you know, these religious discussions, my model is nicer than your model. And and I'm sitting there like, folks, doesn't matter. And Raman was talking a little bit about that one. We see out of the combination between the work from Intellect's Purple Fabric and our knowledge and our experience on the data layers. The extraction, an amazing combination for people.
speaker-0: Awesome. Okay. So, you know, obviously sort of as we move forward then, we have that foundation. I you know, all right, so we have the agents, the agent platform with suppliance pieces sort of ⁓ baked in. You know, one of the immediate sort of questions from an enterprise
Load full transcript
standpoint is hallucinations, right? And again, I don't need to sort of pocket this at length. Everybody sort of understands that these models have a ⁓ you know, tendency at times to, you know, sort of make things up whole cloth, ⁓ you know, sort of answer it correctly and so on. And obviously You know, that's fine in in some contexts if you're, you know, sort of asking for a recipe or something. But, you know, when we're talking about enterprise usage, ⁓ that is very much not fine. So ⁓ Roman, we'll start with you. So what what can you tell us about sort of the the ⁓ you know, how enterprises are mitigating and and ⁓ against hallu ⁓ hallucinations and how worried should they be?
speaker-2: No, I think it's a very legitimate worry. you know, and and and the root cause ⁓ of this, like Boris mentioned, is that it is ⁓ you know, usually organization starts with the model and it's always a model conversation. Model is never the root cause of hallucination. I it is ⁓ it is an actually a data architecture problem. Right? ⁓ when you look at how do you and it's also a sequencing problem. Right? ⁓ don't start with the models. ⁓ it should start really with looking at your foundational intelligence, that is ⁓ you know, what what is your knowledge layer? How do you really ingest your data? How do you apply all of these retrieval strategies? How do you define your chunking strategies? When you're looking at documents which are free-flowing, you know, ⁓ when you're looking at annual reports and sustainability reports and balance sheets, there are flowing documents. You know, you need to have apply different chunking strategies when you're looking at You know, shorter set of documents, ⁓ parent child relationships, ⁓ you know, ⁓ defining what your overlaps are, ⁓ what re-ranker model which you need to apply, ⁓ and and what's what sort of a search capability you need to have. So it's it's actually a sequencing problem, it's also an architectural problem in terms of looking at and solving that. And then ⁓ you know, and then the model d you know does the rest with respect to you know, generating that response on top of what you have really provided to the agent in terms of ⁓ the right level of chance. And and that's ⁓ you know, that's the way to sort of approach it.
speaker-0: Okay. And Boris, how about you? Ma how'd you you know, how worried do you think ⁓ customers should be, you know, sort of about ⁓ hallucinations?
speaker-1: Yeah, ⁓ the interesting part was we had last week in London an event and the statement which came out of that one was it's not not having hallucination back to my model is nice and your model doesn't solve it. The question is the workflows you put behind it, how to deal with and I don't even call it hallucination. You have a ninety percent hit rate of very good answers and results and those can run automatic. How do you deal workflow voice integrated with the other ten percent? Is it More training? Is it more human in the loop? Is it simply excluding and says, please call us back at the bank directly because we want to talk to you? Is it so the workflows behind are actually more critical and that leads us well back to the discussion how do I track and trace my data? How do I have the lineage? Is it a data problem or is it just the SLLM? Let's be very blunt. Not smart enough about the data area, what we're asking for, because it never did an industrial rental agreement. And now we want industrial and not residential. I make this up right now. So those kind of things, that is what's interesting to see. And that's again where the combination comes important. When we are implementing projects, MongoDB delivers to intellect the platform to have the de re-ranker, the embedders, everything in one single flow. And this makes it very easy and transparent from our perspective to deliver the lineage back to intellect. For intellect, it's very easy to see that You have a single lineage. So it's a win win for both sides. But this is the trick what we are doing.
speaker-0: Okay. And so, you know, if if they can get over the initial hurdles with respect to hallucination and then get a ⁓ sort of model ⁓ in agents that they can rely on. You know, one of the things that we see or hear all the time, and again, this is this is ⁓ you know sort of a a common sort of issue, is like, all right, how do we think about challenges around governance, authorization, compliance, you know, sort of et cetera? So first question is does that match your experience and what you're seeing? And second question is, how do you handle it? And so this time we'll we'll start with Boris and we'll come back to you, Rahman. So Boris, go
speaker-1: Abs absolutely. I mean this the governance part and what we call the guardrails are really the important part, which is back to the lineage which I talked earlier about it, to know where data come from is obviously view of the data. But the same as what decision, what act suggested the agent? Why? What were the motivations for the agent and the decision points? And what was the outcome and and annotate this one. These are the kind of things which sound For data people, when you're in the financial service industry or insurance industry, this is normally going for power, but somehow lately the people throw all of that experience out of the window and says, No, Boris, you need to understand it's agentic. And the situation is the same. We still have to solve the same challenges. The governance, the security, the access rights. And my specific piece is who prevents a rogue agent talking to my agent? And we had a long discussion with the financial services company. with Deepak, the CTO of Intellect on that one. And it was very interesting to see. People have not even realized that things like open banking on a gentic level can expose you. You need to have the security. And that's why I like to work with Intellect. Okay.
speaker-0: And Ramon, same question.
speaker-2: Yeah, I mean I
see this ⁓ very much aligned to what Boris mentioned, right? I see this as a five I problem. ⁓ you know, can ⁓ you know, what is the response and can we explain what the response is? Who was really authorized to ⁓ you know, look at that data and ⁓ look at even the response? Where did the response actually ⁓ you know, ⁓ can you point to the right source document? Can you actually point to the right paragraph? To that table from where this information was picked up, especially when when we are working with financial institutions. And and so w when you look at it from Pywise problem, that I think ⁓ translates into how do we really need to look at it from a you know governance standpoint. Second, I think the question also comes in from the data residency and data sovereignty ⁓ perspective. And that's where we, you know, Purple Fabric also provides option in terms of deploying it in their own VPC so that all of that you know. aspects are already taken care of. I I think one more very important aspect what we have been ⁓ you know very consciously and very proactively did is ⁓ you know ⁓ you know we got certified for ISO forty two thousand one, which is an you know AI management governance certification for organizations which ⁓ you know are looking at especially when we work with financial institutions, that's a certification which sort of brings in a lot of trust as well.
speaker-0: Awesome. Okay. So I get two more questions, ⁓ for one for each of you. And we'll start with the the intellect question. So as you think about purple fabric and as you think about sort of the the convergence of ⁓ sort of data data and AI, you know, sort of both now and moving forward, what do you see as the most important challenges and opportunities? ⁓ you know, as these two continue to intersect, ⁓ you know, in the days ahead. ⁓ I
speaker-2: think that's that that's a very interesting question. ⁓ I think working with a lot of finance institutions and enterprises ⁓ for the last ⁓ you know ⁓ couple of years and and also having you know worked with thirty plus institutions where we have actually deployed purple fabric. ⁓ I I think one one learning, Stephen, what what we have really realized is the data platform and AI platform should not be looked at as two different entities. You know? They should be looked at as well interconnected capabilities. And ⁓ you know, the intersection of it is actually where the problem lies. And at that intersection also where the opportunity lies as well. Right. unless you have really solved your data problem, ⁓ unless you have solved your fragmented context problem, until you have solved ⁓ you know, for how do you define your governance in line and not post hoc? in terms of how the data is coming in versus how the data is going to go out and how it is going to interact with the agent tech system. I think those are the aspects which needs to be taken care of upfront and not as a bolt-on or a plug and play ⁓ in in a in a very post hoc manner. This is usually again going back to the sequencing problem. Organizations start with the best of the models, look at ⁓ you know, the more sophisticated agent orchestration aspects. But they really miss out on starting actually from the data aspects, ⁓ and then making sure that both of them are actually talking to each other in a much more seamless manner. So I I I see I see the the challenges, but also from an opportunity perspective, I see there's a lot of opportunity to look at it as one you know, capability under platform, ⁓ where where both of them are talking to each other.
speaker-0: Awesome. That's actually a pretty good setup. So, you know, we'll close out with Boris. So, you know, as we as enterprises think about the relationship between, you know, the database and their AI platforms, you know, again, sort of concerns and opportunities, you know, you see customers in terms of leveraging the data moving forward. And specifically like the role of the operational stack in your view, ⁓ the operational data stack, I should say. So state memory context, et cetera. Like, how does all this come together and how would you advise enterprises to think about it moving forward?
speaker-1: ⁓ my my biggest beef right now, so to speak, is historically AI came out of the analytics space in many of these organizations. So you had ML people who had actually very static data and did an excellent job over those and tried to build some active content out of that one. And when you take a look, what we're talking about right now, as you point out, that's operational data, that's semantic memory, that's real time linkage and behaviour. And from that side You need to have a complete different view to your data. What does real time mean? If I'm online right now, t clacking on my cell phone some KYC and I'm standing already at the branch, you need to integrate this. It doesn't help me to say Boris was last week at the branch and by the way, walked out. We may miss an opportunity. That's great to know. You're laughing about it, but this is an absolutely real incident. So you need to have these real time capabilities integrating that. The real-time data flow is integrated with the real-time decision making and the honesty the positive experience what you want to generate with agenti systems. ⁓ people forget there is a human ⁓ somewhere at the end, not in the process anymore, but at the end there's a customer satisfaction thing. And that means that data becoming actually absolutely v v violently important. And it's not anymore about the data warehouse side. It's really The OTP is back into the game, so to speak. We're back to the nineteen eighties. Old TP calls it.
speaker-0: Fair enough. And on that note, ⁓ I we we shall close it out here. ⁓ so let me take ⁓ the time to thank Boris and Roman for your time. Really appreciate it, and I'll talk to you again soon. Thank you.
speaker-2: Thank you so much.