Specialized AI Models for the Enterprise with Kumo’s Hema Raghavan

Hey everyone. Welcome back here to Techstrong tv. Our next guest is Hamer Rag Havin Hamer is the, uh, head of engineering and co-founder at a company called Kumo.

We can explain all this to you, so stay tuned. First, let's welcome Hamer to our tech strong tv, though show to our tech strong TV show. Hey, Heyer, welcome to Techstrong tv.

It's great to have you on here. Hi, Alan. It's wonderful being here, Alrightyy.

So, as I mentioned, you're head of engineering and a co-founder over at Kumo. We're gonna jump into what Kumo is, but before we do that, let's hear a little bit about your journey. I'm gonna assume if you're head of engineering, you've probably been doing some engineering, uh, and you know, and that you feel strongly about Kumo and, and what the mission there is.

Let, let's, let's hear about your journey. Absolutely. So, uh, Alan, I came to, uh, the United States to do my PhD in, uh, computer science in, uh, a field called information retrieval, which we all know very well now.

I mean, uh, but this was around the time Google and Yahoo and everybody were, you know, getting out there making search, uh, a commonplace. And, uh, my first research paper was actually solving ambiguous queries. And we did that by using a large number of computational linguistic features, right?

So we would say, here's a query, Java, does it mean coffee? Does it mean the programming language? And how do you analyze the text?

And we understood a lot about language. And, uh, I got my PhD actually working in news in the field of news and actually understanding what, you know, how to identify breaking news events in, uh, multilingual, multimodal data. Uh, and, uh, I went from there.

I worked in search ads at Yahoo, and then I went to IBM research working on, uh, uh, this was post Watson, but again, you know, fashion answering systems. And then I eventually came to LinkedIn. So I, this was all, uh, you know, heavy research career.

And then when I came to Lincoln, I was actually hired to do, uh, text because LinkedIn was just starting to get into long form content. And, uh, until then the, the feed was just dominated by job postings and updates and so on. And, uh, I came in and then I realized actually, you know, after six months of trying to pound away at the text features that the network features were the most valuable because, you know, if Alan and I connect right, uh, on LinkedIn after this show, and, uh, uh, you know, that local context of our, you know, uh, social connection is most recent, I wanna know what Alan is talking about.

And that recency of that connection, the edge and the network, makes a large amount of difference. And maybe I, even if Alan talks about three different topics, I won't know what Alan is saying. I want that mantra, right?

So, uh, I learned that network is extremely powerful, and I, uh, moved on from my text expertise to learning about graphs. And actually I led the team that, uh, ran people you may know. I was responsible for all of the notifications that drives LinkedIn's ecosystem and making content relevant in LinkedIn, but through the use of the network, uh, it was in this journey that, and, uh, you know, I I arose to become an engineering leader at LinkedIn.

Um, it was in this journey that, uh, I also met, uh, one of my co-founders, u Kovic, who was a professor at Stanford and the, you know, uh, creators of RAF neural networks. Uh, and, uh, uh, our third co-founder of Kumo is, uh, van Jaki, who it was the CT of Airbnb. We, and I knew him from Yahoo.

And we would meet on occasion and we would discuss, Hey, you know, this, this entire revolution that's happening in text and text was close to me. That's, that, that's how I grew up, right? Mm-hmm.

And, and, uh, but the NEUR networks just learn from raw data. But why isn't that happening to the enterprise data? Why is our ads CTR system at the Airbnbs and Pinterests and LinkedIns still relying on tables and tables of data?

And why are we relying on thousands of engineers to maintains tens of thousands of workflows that do feature engineering? And why can't we have that same revolution that got rid of, you know, those parts of speech that I was human engineering before to solve these problems, like as clicks CCTR, prediction, feed relevance, uh, lead scoring, and how can we make systems more easy to operate and maintainable? And really, it was us, you know, talking in periodically.

We, we were friends, ex-colleagues, uh, that suddenly said, you know, I think there's something, uh, to, you know, uh, create a company here and, uh, let's bring that AI revolution to structured enterprise data. Um, so, so that's, that's what led to Kuma. Unbelievable.

You know, it is a great story because in many ways what you guys were working on, even back then was sort of a precursor, if you will Yeah. To what we call AI today, right? A AI, of course, is not just what we've seen in generative AI in the last two, three years.

My, my friend John Willis actually just, uh, released a book about the history of ai, and, and this is something that's been going on for 40, 50 years, right? Yes. As, as we looked at how, you know, copying how the mind works and neural networks and and and so forth.

So, um, but you're right. What the, the, the, the whole idea of bringing this to private data is, is, I don't wanna say untapped. I think there's a lot of people recognizing that this is, this is an issue and this is what needs to be done.

I, I'll tell you, I saw something recently that caught my eye in that, you know, a, a popular, uh, talking point today is, well, we've scraped the internet dry in terms of gathering data to train the next generation of LLMs to train the next, you know, frontier models. There's just not enough information that we haven't already scraped. But yet, I saw something where, if you look at the amount of data that's kept behind firewalls, Not publicly available on the net necessarily, it dwarfs what's available on the net.

So we are by no means near exhausting the total amount of data that's stored digitally that we could use to train LLMs. The question is, are we gonna be able to, do we wanna have access to that data? Will we be able to get access to that data?

And how do we use that data as well? Or is LLMs is the, the one scientist said maybe a dead end and we need to do world models or, or what have you? Um, I, I don't know.

I, you know, I don't know. I, I would, yeah, Hamer I'll defer to you. You're a lot smarter on this than I am.

I think, uh, uh, what a, a great set of points, right? So I think LMS are, we have to think of them as a corpus of like, you know, uh, like being able to query world knowledge, right? So texts textual knowledge.

But as you are walking through enterprise data, and I see enterprise data every day at as part of kumo, it looks very different. Like, uh, think of a company like GE and enterprise data can in, in involve the tele tree from all of their appliances, because all your appliances are connected to the cloud. That's not text, right?

Then you have your CRM data, you have your ERP data, you have your HR data, and what's relevant for my enterprise is only relevant for my enterprise, right? Um, my CRM data looks so different and is only relevant in the context of my company. Maybe there are other similar companies, uh, you know, B2B SaaS companies in the case of Cuomo.

But, uh, that context is, uh, unique to that enterprise. And you don't probably need to index all of the world's enterprise data to learn a, a, a model. I think each enterprise can have enterprise specialized models for the kind of structure that is relevant to that enterprise, right?

Be, and you need models. Ellips are really next token prediction models, right? They reason over the past and they tell you, uh, the next set of words that you want.

But, uh, the space that kumo is in is actually predictions. So we're not querying even enterprise knowledge. What we are doing is we're actually saying, can you predict the future?

Can you actually say, uh, I'm ge and do I know that this appliance is going to break down? Uh, I'm DoorDash and do I know that this restaurant is relevant for this user? I'm Reddit.

Do I know this ad is relevant for this user? And Snowflake? Do I know that this, uh, account of mine is gonna churn in the next quarter, or this account is going to upsell in the next quarter?

So those are the kinds of customers. Those are the kind, that's the kind of reasoning that umo brings in. It's predictive, uh, you know, it's predictive capabilities.

And yes, I think the fa the, the space is evolving. So yes, you might need some other kind of models outside of l LMS for reasoning, uh, and so on. And I think, uh, you know, uh, we'll come to a place where there'll be, you know, few different options.

Excellent. Yeah. If you don't mind, I, I wanna focus in on Kumo a little bit here, though.

So an amazing founding team, it sounds like, right? Just three literally PhD doctors, you know, Stanford University professor, fantastic accolades and private industry coming together to, to tackle a real problem. Talk to us about founding the company, raising money, coming to market, building out the rest of the team.

How's your experience, Becky? Wow. Uh, it, it's, it's, uh, a lot.

Uh, it, it's fun. It's a journey, right? Like, especially going from big tech to founding a company.

Um, I think, uh, the, uh, when we started out, we were, again, you know, I, I think founding good founding stories come from a pain that you've experienced yourself that you wanna solve, right? Like, that's where that passion comes in. And I think the fact that we had spent a lot of time in consumer and big tech, and actually dealt with building these large scale recommender, uh, kind of systems and, uh, uh, how painful they were to build and scale out was what led to the journey of Cuomo.

Um, uh, I think, uh, uh, you know, engineering came naturally for all three of us, because that's what we did, right? Um, I think definitely, uh, it's different when you're a younger company. You, you've got to move faster, test hypothesis faster, learn from customers much faster.

And as a founding team, we learned a, to actually listen to our users, uh, respond to feedback, we learned how to, oh, so Kumu, uh, actually prices based on value, you know? So in, uh, so, you know, just understanding what value meant, right? How do we drive millions of dollars of impact and sometimes billions of dollars of impact for our customers, and, uh, how do we draw long-term value?

Like these were, these were all part of back learning curve. And of course, fundraising is, it's, uh, you know, uh, own journey and, uh, yeah. Yes, yes, I've been there, done that.

Yes. I, so before I started Techron, I, I did four or five venture backed startups, and, um, yes, I know we're well familiar, um, but you wanna know the truth, once you get it in your blood, it's like a fever, a virus, right? Yes.

It's hard to get rid of as well. Um, I won't do anything else. Exactly.

Exactly. Yeah. I, I gotta ask you a question though.

So you guys are really focused on predictive AI for structured data. What about unstructured data? Do we ever get our heads around that?

Oh, yes, absolutely. Right. So, uh, any enterprise is a mix of structured and unstructured data, right?

So if you are a retailer, uh, you have your product catalog and your product catalog, uh, maybe items that are, you know, available on the public web. So an LLM can actually tell you that, uh, this uh, uh, um, brand of, uh, toothpaste is very similar to this other brand of toothpaste and maybe pri a, you know, price or in other similarities dimensions, right? Like prop, like properties or, or some organic brand is similar to another organic brand and so on, so on.

There's definitely, uh, knowledge, world knowledge that an enterprise can learn from in its own structured data. And Kumu actually lets you bring that kind of information into its models, okay? But through what are called embeddings, or you can think of them as these vectors that tell you that, you know, item A is similar to item B, but what kumo does on top of that is actually a reason over, you know, did all the users who bought this brand of toothpaste also engage with this other product?

Uh, maybe the toothbrushes from the same brand, right? And then draw the kind of relationships that, uh, uh, based on the transaction activity that you might see within that enterprise, right? Price sensitivity of its users, the, uh, demographics of its users delivery times.

Like there's a lot of a structured information inside the enterprise that Como will reason about, but you can actually bring this other catalog, world knowledge kind of information into Cuomo. I love it. You know, I don't know if, did we mention the website U rl?

Not yet. Not yet. It's Omo, KUMO Do ai, Ai.

Okay. That was easy enough. Last topic I got for you, hammer, and that's AI agents.

I was talking to someone the other day, and you know, as I get older, I guess I become more sensitive to this stuff, but I was talking to someone the other day and they said, oh, generative AI's old and over already. It's all about agent ai. And I thought, Hmm.

It's been just, it was just 2, 2, 3 years at most, and now it's over. And we're all about the AI agents, and I don't necessarily disagree that agen is huge. I just, you know, as I think we were talking off camera, sir, another survey, most enterprises said they have 10 or more agents working right now, digital workers again Hmm.

Fact point, right? Then I look at my own world and I consider myself a tech kind of guy, and I'm having a hard time making my AI agents do the task autonomously that I want them to do. And, um, I'm wondering, are we still like in this future state talking about what it will be and when does it become real?

Or am I just missing the boat, Alan? I think, uh, we are in the same boat because, uh, okay. We've been playing around with agents, uh, ourselves at Comar because, you know, uh, we, we're a lean team.

We want, we want to, uh, you know, uh, grow our team very intentionally diligently, and, you know, have high horsepower individuals who are not bo down by mundane tasks, right? Um, I do think, uh, I'll tell you where I see very basic impact, like, you know, uh, uh, very simple things like, uh, routing my SRE tasks to the right team, uh mm-hmm. It, I think it kind of works, uh, but it's largely role-based.

And where I see that and, and it works for a team the size of Como, so it's, it's mostly automation at this point. It's not, uh, uh, you know, making, but Automation is not autonomous. Exactly.

Exactly. Automation sounds like BPO to me. Exactly.

Or BPA, excuse me, business process automation. Exactly. If this, then that.

Yes. So I think one would hope that the, if then the, that the, and that is then replaced with some predictive system. It could be, uh, a kuo like system that actually says, you know what if like, I learned those rules, right, based on patterns that I observed.

So agents have, so what, what makes us fundamentally human, right? Our brains are capable not only of all the knowledge that we amassed over the la you know, the, uh, over the years lifetime of our lifetime, lifetime, but also this ability to abstract that into forward looking decisions and based on past experiences also informing that. But this human brain has this ab ability to do the coulda, woulda, shoulda of what should happen.

Mm-hmm. And I think Kumu wants to complete that other half of the brain, but I think agents are missing, you know, uh, that half. And, uh, yeah.

I, I, I, I think until then, it's just glorified business process automation. Yeah. The other aspect, Alan, is agent security and access control.

I know this is close to your heart, but, uh, yeah. You know, I just wanted to say that's something that's gotta be figured out. Yeah.

No, it's, right now it's an oxymoron, right. Even mention it in the same breath, but, but here's, here's where I guess I'm the boy who goes in the room full of horse manure and says, I know there's a pony here. Yeah.

Um, right. I do think it offers us the promise of figuring out, I think we will figure out security. Yeah.

It may not happen as fast as we want, but eventually we will. But I do think it, it has the potential to get that other half, but like everything else that I've seen in my lifetime, we tend to wanna rush these things. Right.

You know, the old saying, you can't make wine before it's time. Right. And, and we've gotta give these things a chance to play out, but we're always onto the next thing in tech, right?

And, and this may be a case where it's like, I was talking to a quantum guy about when is Q day happening? When's q mm-hmm. You know what, the day it happens, you may not know it, you may not know it till six months after, but eventually yes.

Kind of the, the market, if you will, catches up to the leading edge. And I think that's what we're gonna see where the chat to as well. Absolutely.

Well said. And I think with all such technologies, whether it was the internet, whether it was Yeah. Mobile, you know, there's that, we don't know what's gonna happen, this mad rush, but over time it has changed our lives and it's a part or drastic drastically.

Yeah. My children don't know a world without the internet. It's, it's utility for them.

Absolutely. So, And their children will never know a world where people drive cars, cars drive themselves. Yes.

There you go. Yes. Anyway, Hamer, thank you for coming on Textron tv.

It's been a pleasure. Best of luck with Kuo. Thank you.

Do come back and keep us posted on, this is a really, look, is a, this is a, a subject that we cover very closely anyway, and I'm personally interested in, so I'd love to hear more. Thank you so much, Alan. It was a pleasure being here.

And look forward to, uh, you know, working with you. Absolutely. Thank you.

Hey, Raghavan, head of engineering co-founder here at Kumo on text, on tv. We're gonna take a break. We'll be back with more.