Transcript
Mike Matchett: Hi, Mike Matchett with Small World Big Data. And we're here today talking, of course, about AI. Everything's AI, but it's getting out of hand for a lot of people. I have a lot of AI proliferation. We have things like shadow AI, we have cost overruns with AI, we have compliance issues with AI. This AI application proliferation is taking over in a lot of companies, and we can add agents to that and identify what do you do about it. So we have today an answer for you. So just hold on and we'll get right into it. Hey welcome to our show. Um, did I get that right? Yep. Yeah. Good. Uh, you are a machine learning expert, and you've got a bunch of machine learning experts over there. I am a big fan of machine learning, but everything's called AI today. Maybe you could just start by telling us what your sort of machine learning background is. And a little bit of philosophy about machine learning going into this AI age and how, how you can really maximize the value of AI by using AI. Poonacha Kongetira, CEO, Classie AI: Sounds great. Thank you so much for having me on, Mike. I appreciate it. And I'll just say that, you know, my sort of like journey in this field sort of began began about ten years ago, I would say. Um, I joined Google. I was fortunate enough to contribute to, uh, the TPU program. The TPU is the accelerator that Google uses. It's what Gemini, the model is turned on, trained on. And besides that, you know, um, did a bunch of other infrastructure work which involved hardware software co-design in Google as well for things that power YouTube, things that power, you know, then networking machinery, etc.. And, um, my co-founders in this, in this endeavor of ours, um, you know, one is a professor at Stanford. His name is, uh, most recently he's also, um, co-founded a company called Sambanova, which is I first, you know, soup to nuts startup. And, um, among my co-founders, one of whom is his student. Another one is, uh, an AI researcher who's been involved with, um, machine learning models, which were all the way back from random forest, you know, to, to, to present day, you know, um, multi-agent, uh, you know, frontier, uh, agent ecosystems. So, um, so I think that's, that's our background. And we, I think what, uh, what two years ago, we, we, we sort of like put a pin in the graph of AI adoption and the shape that these systems would actually take. And we said that, you know, um, at the rate at which models are improving and at the rate at which, um, applications are beginning to use models soon after the ChatGPT moment in 2022. Right. It's shortly going to be the case that um, foreign enterprise, um, an AI application which uses an embedded local model or a frontier model, or an open source model is going to be present on every surface, whether it's your browser, whether it's your endpoint, whether it's your Kubernetes surface, whether you're interacting with a previously SaaS workflow, which now has embedded AI, it's going to be everywhere. And so our thesis was that in this brave new world, um, you needed something a little different. And there's something different, right? Comes from our observation that the underlying math, right, which um, large language models and every other sort of model uses is necessarily, you know, goes through the stochastic and probabilistic process. And therefore the, the behavior can be somewhat variable. And to add to this, um, you know, most application, 99% of all application builders out there are going to use a model that somebody else has trained, which is a black box to them. Right. And one of the central ways these things are different from traditional software is that a, they're black box is the variable behavior. And they can sort of spontaneously, um, come up with new workflows in response to a prompt. Right. We have never had this kind of behavior before. Mike Matchett: And so it's just making it up as it goes along really. And by the way, on your randomness and probability thing, I people say, no, no, it's really smart. And I say, look up softmax as an algorithm sometime and tell me, tell me, come back and it's talk after that. But uh, yeah, yeah. So, uh, it definitely gets complex. So what, what becomes, what becomes the major challenges once you get to this, uh, AI proliferation. Poonacha Kongetira, CEO, Classie AI: I think first thing is you've got to be able to detect where they are, right? Without knowing what's running on your surface. You actually, um, are at risk. And I think, you know, every enterprise has to sort of think about themselves as, um, whether they know it or not. Um, um, you know, AI models are running on their surface and in some cases exfiltrating data. So the thing you have to think about is how do they maintain the trust of their customers who have signed an agreement with them, which says privacy operating policy. You know, secrets, whatever. Right? How do they maintain that trust? And secondly, as an enterprise themselves, who becomes a consumer of AI, if not now, tomorrow, right. How do they manage the metering? And do they have an opinion on how to use AI cost effectively? Because all applications, you know, with embedded, uh, you know, uh, LLM calls and stuff like that are going to have this problem. So do they have that opinion? Can they manage their spend? And can they, can they keep their customer trust? I think those two things are fundamental to every enterprise. The way we help solve that problem, that's the problem that we have to solve is that we have this thing that we call, um, you know, supervisor AI using an operational confidence layer. And we have some attributes of that layer which say, this is how you do it. And now you can actually proceed forward. Um, you know, on your AI journey with confidence. Mike Matchett: All right. So let me, let me just recap that a little bit. So, you know, I pulled out two words, right? It's trust and efficacy or efficiency. I'm not sure which one of those is really the thing there. But really it's about cost control at the end of the day or cost optimization, I'm spending making an investment and I want to get the maximum return on that investment into my. I think right now people have very little control or handle on what they're doing. They just are saying, hey, go do it. And then getting surprised. Right. So that's a big problem. And the trust issue is just going to keep coming back around. I mean, we don't we don't really know. I, I do, I do think that the, the problem with this AI swarming proliferation is that you put AI into everything, as you were pointing out. So when you when you take on that first layer of discover, like even just and then analyze because I know that's where you're going to go next. Uh, how do you get data from every surface that's out there that's using AI? Poonacha Kongetira, CEO, Classie AI: Well, it's really interesting. So I think first, you know, there's three principal surfaces on which, um, you know, AI applications get consumed, right? Um, either, uh, enterprise workers are interacting with their applications through their browser, or they've downloaded some version of open cloud for work or codex or most popularly coding agents like cloud. And so you want to know what's going on. So those are two primary surfaces. Um, the browser also tells you when you log into a SaaS application, whether you actually have embedded AI there, right? Chat bots, stuff that you can, you know, maybe paste stuff into which you're not sure about. Um, and then finally, you know, most enterprises are deploying, um, um, you know, applications that they're building themselves on some sort of Kubernetes framework. Now they could be using Docker or this or that or bare metal, but Kubernetes is the, is the main main one if. And so, uh, we have sensors, you know, which act as Kubernetes operators. The nice thing is because of general convergence of design patterns in this space, right? We're able to using, uh, you know, this handful of sensors be able to, um, cover this very vast surface, you know, everything from the endpoint to core enterprise operations. And, um, basically what we're able to do is first, you know, see when a container is spun up with a particular application, um, or, um, a user, you know, downloads and onboards and brings up a new Harness, you know, on their endpoint, we start creating an inventory. So the discover phase actually goes into that, start creating the inventory, figure out what, uh, what the applications are connected to, what credentials are being transferred from one user to the other. These are all like massive failure points, right? And so we, so we see all of this stuff. And then we, as you said, get into the analyze phase where essentially, you know, to really know what's going on, right? Really, really you have to be able to, uh, you know, have this, you know, multi surface transcript of activity, which serves as your AI record for a single question that you ask it, right? That explodes into a thousand different things and you have to be able to like grab that and bring it back and put it together in real time. Mike Matchett: Real time is a challenge, right? Poonacha Kongetira, CEO, Classie AI: That's the that's the thing dynamic. Mike Matchett: When you have agents of agents and agents swarms, it's going to be it's making it up as it goes along. Right. There's no pre there's no pre testing what it's what's going to happen. Poonacha Kongetira, CEO, Classie AI: No. And and I like love to use this jigsaw puzzle thing right. Because putting this thing together as a jigsaw puzzle. And um, if you find pieces of the jigsaw puzzle and throw it into a box, it's not going to assemble itself. Right. So, so, so you have to like, you know, tag and bag these puzzle pieces as they're coming in so that you assemble it in real time. And as the record is building up, you're like throwing up insights, you're blocking behavior, you're doing things like that, right? Alerting. Mike Matchett: So that's yeah, I mean, I mean, some people walk away with the vision of putting your pieces of jigsaw puzzle in the box and shaking it and hoping open it up and find a puzzle complete is not going to happen. No, it's going to be more like DNA sequencing, where you got to take all the little bits and match them all up, and it's going to take take a while. Um, yeah, just just philosophically and I hate to do this because there's a lot of deep technology here, but are you thinking this is more of a cybersecurity attractive solution or a compliance solution or a cost optimization solution in the IT audience, for example, should be paying paying most attention to what you guys are doing here at Classie. Poonacha Kongetira, CEO, Classie AI: The first stage that every organization has to go through is discover right, discover what's going on. Right? And that, I think is something that cybersecurity people really care about, right? Once you deploy the sensors and you start forming your inventory and whitelisting and blacklisting things so that you can control what runs, then you can decide how to optimize its usage, for which I think the audience may expand from cybersecurity to more, you know, CIO, how do you control your spend, you know, compliance? Are we doing things which are like proper, right? Um, and, uh, you know, subservient to the policy that we have as an organization. And how do you actually then set up your policies in a way to govern it? Right. So I think it's sort of like lands in cybersecurity at the discover phase and expands beyond IT to compliance and CIO. Mike Matchett: Right. Because because I, you know, you're, you're, you're very close to like a lot of analogous things that I've covered and worked on in my life, even though this is a whole new territory. But we talk about observability and systems management. We talk about things like, how do you collect data from various points? Do you centralize that data? Do you process it? And you have some very unique and scalable answers to all those things. Now, just can you tell us just briefly about the scalability challenge now with AI and how you're more uniquely built to handle that than, say, just like your traditional network security kinds of tools? Poonacha Kongetira, CEO, Classie AI: Absolutely. I think the scalability problem in AI stems from like two things. I think one is your ability to handle the surface and put the information together and to take action in real time. I think that's agents are far the harder, faster than anything else. You need to take action in real time to enable you to take action in real time? The machinery that allows you to analyze and then take action cannot be so expensive that you can't afford. You have to make it shouldn't. It shouldn't be so expensive that you have to like, think about it, right? It's got to be like 5 or 10% of the application cost to build that. What we've done is built a machinery with, um, with owned adapted models that can run locally to your virtual private cloud. You can deploy it either use a SaaS tenant, which is single tenant or deploy locally and run it on, on your premises. And the nice thing with this is your data stays local. Mike Matchett: You maintain that sovereignty and that compliance. Yeah. Right. Poonacha Kongetira, CEO, Classie AI: And then the, um, and then in the supervised tier, which is like a next quarter product, right? We, we want to allow you to adapt these models to your usage. And so when you adapt those models, that, that means, you know, effectively, you know, some version of training, right? Or self-learning. And at that point, the models are yours. Like we don't. Mike Matchett: Take. Poonacha Kongetira, CEO, Classie AI: Them. Okay. Mike Matchett: You're not sharing. You're not sharing my embedded IP of how my organization works and secures things with everyone else. Yeah. Okay. Poonacha Kongetira, CEO, Classie AI: Yeah. We just took a look at, you know, like, how would we want to consume this stuff? And we're like, yep, all this infrastructure has to be customer owned, customer operated as and when you're ready to consume it. And we have a journey that you can take with us. Right? Mike Matchett: And, and I'm going to say just even with that, that sort of little nugget thrown out on the table, if people are really paying attention by having a hybrid kind of supervisory regime observability, discovery analysis, and also an ability to supervise and go in and actually do active control, but done so that it preserves sovereignty of the data collected and the models trained does require an awful lot of deep technology that we have not scratched the surface of in the last ten minutes. Right. So I would encourage everyone who's looking at this to sort of read between the lines and say like, wait a minute, there's a lot more going on here to make this sort of higher level conversation sound smooth, right? We get trust that we get cost control, but there's a lot that's going to go on behind this. As you guys develop. So, um, we're sort of running out of time here. But before I even ask you what to look for, if someone is deploying AI and AI agents, what would you say should be their first concern as a sort of a best practice recommendation? Maybe even outside of what you guys do, what should be the, the first thing or two that they focus on? Uh, if they're trying to get control of this from, let's say, a cybersecurity perspective. Poonacha Kongetira, CEO, Classie AI: I think the, I think the first, most important thing is, you know, there's been a rush to deploy stuff, right? And, uh, in many ways, um, uh, you have to have that rush to deploy so that there's adoption. But, um, you know, traditional governance practices sort of get pushed to the side, which is okay when you're trying to get some adoption, but you know, that journey can't, you can't go too far. Mhm. On that journey without actually thinking about governance. And so setting up some notion of an org structure which says, this is how this is the heads I need to bring together, right? To actually build up a policy framework for my organization. And how do I now implement that policy framework? Right? What is what is the principles of tooling that we actually need to bring together so that this new beast with variable behavior and black box characteristics, right, can actually be harnessed, to use a popular word, right? Is, is, is really the challenge that every organization should take up. The one more thing I'd say is you have to get started, right? I think you have to get started. Um, you can always keep analyzing and coming up with requirements. But the fact of the matter is this thing moves faster than, faster than your requirements list. So the important thing I think is to, um, you know, you know, choose a provider and get started on the journey, right? And so you can evolve as you go. Mike Matchett: So, so I'm just going to summarize that. The first part of what you said here is no matter the technology involved, it's always a people problem at first. You need to solve that people problem by getting the right people together to do that. That was a wise piece of consulting advice. Uh, that sort of comes up over and over again. Yeah, take that to your board. It's always a people problem when they ask. It's always a, it's always a people, um, at first. Okay, then. Um, okay, so we're kind of running out of time here. If someone wants to dive more into what you're doing, they're Classie. Uh, where would you point them at? Obviously you have a website. Is there some particular thing that someone would do to do maybe a little more research or. Yeah. So on this. Poonacha Kongetira, CEO, Classie AI: Yeah. So there's so our website is dub dub dub classie.ai. That's Classie, spelled with an I. And if you go there, you'll actually see what our product is about. There's a on the resources page. There's a white paper which I actually co-wrote with, uh, Roland Cloutier, who's, um, former CISO of TikTok, um, ADP, EMC, a bunch of, you know, things, right? It's called, uh, the illusion of control. And, uh, you know, I think that sort of sets up, sets up some of these best practices that, you know, any organization may want to adopt, whether you work with us or not. Right. And then certainly, if you want to work with us, there's a there's a place there to click in and, and, and get in touch with us. And we'd be delighted to sort of, um, work with you on your AI journey. Mike Matchett: I know it's early days yet for Classie in some ways, but it's a fast moving market and you guys are fast developing solutions to help manage it and, and supervise agents and agents and agent networks and agent swarms and all that stuff. So thank you. Because I think I think the world needs this, uh, more than ever. Uh, every month I'm hearing more and more of the challenges of things getting out of hand for people. So I think this is a good step in the right direction. Um, check it out. Classie. I thank you for being here today. Poonacha Kongetira, CEO, Classie AI: Thank you so much, Mike, for having me. Been a wonderful conversation. Mike Matchett: All right. Take care, folks. Poonacha Kongetira, CEO, Classie AI: All right. Bye.
