Applied Compute's Yash Patil is building continual learning AI agents that absorb institutional knowledge

and HR built to evolve with modern small and mediumsiz businesses. And without further ado, let's continue our lightning round and bring in Yash Patel from Applied Compute. How you doing?

Well, what's going on?

Good. How are you?

Good to see you again.

Good to see you. Since this is the first time on the show, long overdue, uh, please introduce yourself and the company.

Yeah. Yeah. So, so my name is Yash. I'm the the CEO of Applied Compute and what we do is we build what we call specific intelligence for enterprise.

Um, and what that means is, right, like you know, AI is is sort of um, you know, going at breakneck speeds. These general models are getting better and better week over week. Um, but you know, if you're using the general thing, you're kind of never um having your competitive edge. So what we think is there's a ton of latent knowledge or subject matter expertise that's kind of in an enterprise and what we want to do is help enterprises capture that imbue that into their agentic workforce and you know start to to scale up their their their agentic co-workers.

What is an actual onboarding process look like for an enterprise? Is it just sort of turning loose?

Before you answer, I just got to say I love the I love specific intelligence. I don't I've got enough general intelligence. I really like some specific

I think a lot of people are feeling that they're like everything's at 99% but I want 99.999%.

Exactly.

Yeah. Yeah. I mean like um you know there was that MIT gen state of genai paper that came out a while ago and you know the the thing it highlighted you know the reason 95% of these AI pilots fail is because these things don't really do the last mile. So they don't adapt to feedback. they don't get better the more you use them kind of like an employee would and um you know that's really what you need in enterprise in order to actually have like a productive employee um you know you can do kind of simple automations we've seen a lot of like workflow building um and that's great so I I kind of think about this as like RPA plus right you had um

you know you used to have like sort of like your click and drag RPA now you're introducing models into it but to do sort of like real cognitive work that like knowledge workers are doing, you kind of need to go a step above. A lot of that is context, but a lot of what we do is is fundamental research on the model level, too, because we think um you know, sort of building this uh this next agentic employee is going to require um sort of touching the entire stack

and and what does your stack and supply chain look like? Are you a beneficiary of open source? Are you partnering with big labs? you obviously have a lot of experience with Big Labs, but uh what does your what does actually deploying one of your products look like?

Yeah, so so so we're a platform. So we deploy a platform inside of an enterprise. Um we really want to sell the entire stack. So that's um you know being able to plug into all of your systems of record and and data because you know we think context is there. It's just really fragmented right across all these different applications and even people. A lot of stuff is in people's heads. this this tacid knowledge. Um so you know we plug into all of your data. We have our RL proprietary RL post training stack where we can train these sort of like um reasoning models directly on top of your data. All the infrastructure around models which makes them agents right so uh there's the model but an agent is really like where's that model running what tools does it have access to all the permissioning and authentication. Um and then above that is like the application layers. So how are humans interacting with these agents? um how are they like sort of guiding it, instructing it on what to do and then the observability around of all all of this um and what we think our value really is is is closing the loop. So, um I was, you know, everyone's been talking about continual learning. Um I think that's entirely how this space is going to go, right? Like we have offline evals today. Um that's because that's kind of the best thing we have to benchmark. Really, these models are getting so damn good that um you're going to be evaluating them based off the real work that they're doing um in the enterprise and like while you're actually tasking them with stuff. So, um, we basically help capture all of that information, um, turn it back into, uh, context and data that we can use to continually train these things.

Uh, I don't know how much you can share or how many of how much is this is a secret sauce that, uh, that that maybe you'll talk about on a podcast 3 years from now, uh, when when you've already won, but uh, what how do you how do you get all the context that lives in people's head uh, into your into your system so that it can be used across the organization? is like I imagine you've tried a bunch of different attempts and and ways.

Yeah. Yeah. So, so I think there's like a couple of of um you know standard sort of recipes that you can use to to start these things. And I actually want to be super clear. I think RL is is you know sort of one of the best ways to train these models today. But it's not going to be the only thing. And it's going to constantly evolve. It's it's honestly quite uh nent in its stage, right? You know, I think Karpathy and a lot of these other folks have talked about like how simp overly simplistic it is, but um you know, a good example, Jordy, might be uh embodied work, right? So, humans go and spend a ton of time going and creating artifacts that they spend a lot of reasoning effort on. Um and what you can actually do is you can look at these artifacts, these these final outputs, and sort of say, hey, this is what good looks like. And then optimize models, sort of train models to produce things that look like that. So um you know I think a a recent example of this uh that we actually you know we put out some some uh collaborative research with Meror like 2 days ago um uh sort of hill climbing on this this new agentic benchmark they call Apex so sort of a professional services benchmark across law investment banking consulting management um and uh yeah we're we're sort of number one on the corporate law subdomain there I think uh got bumped down to yeah and like number four or five overall. So, um yeah, it's crazy how much you can do with with these the small amount of data.

Where where where is applied compute like where what what sectors are your customers like having their mind blown in the way that software engineers have generally with with codegen uh tools over the last year?

Yeah, it's it's a great question. So, so we're really targeting um you know institutions where there's a lot of sort of builtup knowledge and context uh over over decades right so this is like financial services insurance healthcare bio um places where data really really matters um I think you know even even the coding domain right it's you know there's so much uh there's such a high ceiling there in terms of the types of of of models you can go and train so you know we we've been doing some some work with with cognition helping train some custom models there. But um yeah, I think like places where there's a lot of sort of like institutional knowledge that's where where this sort of imbuing it into these models shines the most.

Mhm.

Well, thank you so much for coming on the show and

do we have to I think we got to hit the gong.

Oh yeah. Uh yeah, I mean the the fundraising it was it was a little bit ago, but how much did you raise? We want to hit the gong.

Yeah. So so we raised uh 80 million last year. Better late than never.

Uh great great to finally have you on the show. Uh you're welcome anytime.

Yeah, we'll talk soon.

I hope we can hang again soon.

Have a great rest of your day. We'll talk to you soon.

Let me tell you about Crowd Strike. Your business is AI. Their business is securing it. Crowd Strike secures AI and stops breaches. And let me also tell you about Railway. Railway is the all-in-one intelligent cloud provider. Use your favorite agent to deploy web apps, servers, databases, and more. While Railway automatically takes care of scaling, monitoring, and security

get