Founder Chats - Daulet Amirkhanov - Code Story: Insights from Startup Tech Leaders

[SPEAKER_00]: Hello listeners, today we are dropping another episode in our chat series, specifically on the founder side, hearing from those scaling the companies themselves.

[SPEAKER_00]: In this episode, we are talking with Dalat Al-Marcanoff, founding engineer of BDI.

[SPEAKER_00]: Dalat is going to take us through his years at Meta and Cogny, leading into how he's building BDI to take on compliance audits in AI automation.

[SPEAKER_00]: Dalat, thank you for being on the show today.

[SPEAKER_00]: Thank you for being on code story.

[SPEAKER_01]: Thanks for having me.

[SPEAKER_00]: Absolutely, really excited to dive into all of your experience into BDI, the things that you've taken away from the work that you've done.

[SPEAKER_00]: You've gone through a few years of high throughput, reliability infrastructure at meta, right?

[SPEAKER_00]: To engineering the graph rag engine and semantic memory systems at Cogny, and now you're founding engineer at BDI.

[SPEAKER_00]: So you've run the gamut, you've done some big things, you built some big things, and

[SPEAKER_00]: BDIs and Andreys and Horowitz back start-up building autonomous agent infrastructure for compliance audits.

[SPEAKER_00]: Okay, so here's the question, how did that journey shape the way that you think about engineering for the age of autonomous systems?

[SPEAKER_01]: Early only my career, as you mentioned, I worked on a high throughput reliability infrastructure at Meta.

[SPEAKER_01]: and that early experience has helped me to build the foundations of me myself as a software engineer as well as a problem-solid in general.

[SPEAKER_01]: In retrospect, I'm really glad for the way that I worked first at NETA because at the right after NETA, I think with the proliferation of autonomous systems, agente systems, it has helped me to leverage my experience that I've gained early on in my career.

[SPEAKER_01]: and make me just much more efficient software easier by leveraging autonomous systems.

[SPEAKER_01]: And I think you have shaped me as like a early on, as a problem server and as solid software easier.

[SPEAKER_01]: But now, as a problem solver with these really awesome, agentic and autonomous systems that might disposal to be a several X more efficient problem server.

[SPEAKER_00]: Absolutely.

[SPEAKER_00]: Now, that makes total sense.

[SPEAKER_00]: And, you know, it's empowering a lot.

[SPEAKER_00]: It sounds like it's empowered you a lot in your development life cycle.

[SPEAKER_00]: Before we dive in a little bit deeper into that, tell me about she maybe outside of technology.

[SPEAKER_00]: What do you do for fun?

[SPEAKER_01]: Outside of technology, whenever I can, I go up, take my cycle and go out of the city, as far as I can.

[SPEAKER_01]: Of course, a couple of times a year when I can, I'm really avid, a fan of sking, and I also do run whenever I can, but lately I've just been fully committed to cycling, but I think other than that, yeah, it's been just coding and cycling, I think, that's how you could summarize most of the time that I spent.

[SPEAKER_00]: Okay, let's dive into a little bit deeper into the meta years, right?

[SPEAKER_00]: Let's zoom into it a bit.

[SPEAKER_00]: The meta's obviously a big scale company and for those that haven't worked at that type of scale, right?

[SPEAKER_00]: What was the exact piece of logging and reliability infrastructure that you owned?

[SPEAKER_00]: And you mentioned this word earlier, high throughput.

[SPEAKER_00]: What does high throughput actually mean in numbers there?

[SPEAKER_00]: And I think to take that a step further,

[SPEAKER_00]: What's one specific architectural decision from those years that has been found national and how you build software here in today?

[SPEAKER_01]: So let me start from first one.

[SPEAKER_01]: I've been working, as you mentioned, at Meta, like working on high throughput systems.

[SPEAKER_01]: And what I mean by high throughput systems is I could probably start by saying that I've

[SPEAKER_01]: Most of the time that I worked at midday, I worked in ads.

[SPEAKER_01]: So it's the backbone of the company financially, but also it's the part of the company that's faces the most traffic.

[SPEAKER_01]: So when people say billions of active users, those billions usually hit first the ads in the infrastructure.

[SPEAKER_01]: And the part that I was working on was part of the ads.

[SPEAKER_01]: And it was only a fraction of the overall traffic, but still it was at the smallest, it was millions of log hitting our systems daily.

[SPEAKER_01]: And that's just a tiny fraction of the overall ads product.

[SPEAKER_01]: Specifically, I was working in a part of the ads that's called Shops Ads.

[SPEAKER_01]: It's an internal shop framework shop product.

[SPEAKER_01]: That is currently available for US, like customers, like users, and I was working specifically on logging.

[SPEAKER_01]: I did, like I did both logging and reliability, but most of the time I worked on logging.

[SPEAKER_01]: I worked at Meta for a bit over three years and I think I've worked on so many interesting projects.

[SPEAKER_01]: The thing that still shapes me as both as a software engineer, but as a problem-solid general, is at that scale and with that, like at that scale and without throughput, what you want to do at Meta is...

[SPEAKER_01]: move fast, ship things fast, and check my ship as small as fast as you can.

[SPEAKER_01]: So that kind of mindset of chunking your problem into smallest, digestible pieces that you can ship as fast and check your hypothesis as fast with that stuck with me ever since.

[SPEAKER_01]: And I think it's kind of like one of those, one of those things that still shapes me the way I approach.

[SPEAKER_01]: things, especially as a software engineer to this day.

[SPEAKER_01]: And I think that's kind of one thing that I like really took away from those years.

[SPEAKER_00]: Yeah, sure.

[SPEAKER_00]: That's core to software development and shipping the essentially smaller commits right or shipping smaller pieces that you can move more agile.

[SPEAKER_00]: That makes total sense to me.

[SPEAKER_00]: A lot of infrastructure engineers, they stay in for a right.

[SPEAKER_00]: And you made a deliberate move from human scale systems at

[SPEAKER_00]: to agent scale systems at Cogny, right?

[SPEAKER_00]: So what did you see in that moment that convinced you, AI agent infrastructure was the next distributed systems frontier and not just the current hype cycle.

[SPEAKER_00]: I'm really interested to hear what you have to say there.

[SPEAKER_01]: So as I mentioned, I worked in the interest structure for most of the time that I was at meta, but just before I just before I make the switch, I internally made a decision when I was working as meta to work and pursue my interests on ML systems,

[SPEAKER_01]: at ads.

[SPEAKER_01]: And that was my first step that I took because I was really interested in how to personalize a personalized recommendation systems work.

[SPEAKER_01]: How do you leverage that?

[SPEAKER_01]: And I think that was my first kind of like step in that direction.

[SPEAKER_01]: And while working there, I was I got more more interested in AI in general.

[SPEAKER_01]: And around that time, there was a lot of

[SPEAKER_01]: interesting things developing the alms were already big thing back then, but like small things like retreatal of method generation have been already picking up and I've been really falling that closely and around that time I was like it's been a bit like around three years that I was at meta and I think I took a lot like a grew lot like personally and as a software engineer at meta

[SPEAKER_01]: But I think that's when I start feeling this need of, I really want to switch gears again.

[SPEAKER_01]: I'm trying to learn how things are actually moving beyond big tech.

[SPEAKER_01]: So that's when I made this decision of, yeah, I want to try to see how things are out there.

[SPEAKER_01]: And that's when I came across Cogny.

[SPEAKER_01]: And that's how I could make this switch.

[SPEAKER_01]: I think like when I was talking to people at Cognitive, one of my first calls was with PhD scientists that were actually part of the team.

[SPEAKER_01]: And that was the final thing that sold me.

[SPEAKER_01]: Yeah, I want to try this, like blank story short, worked on Cognitive, and fast worked here.

[SPEAKER_01]: I am working on agentic systems at Beat.

[SPEAKER_00]: Amazing.

[SPEAKER_00]: That's super cool and Codney seems really interesting.

[SPEAKER_00]: I'm sure talking to those folks.

[SPEAKER_00]: He sealed the deal for you.

[SPEAKER_00]: Codney being the agent infrastructure layer, right?

[SPEAKER_00]: It's a graph rag and semantic memory company.

[SPEAKER_00]: And your work there was on the agent infrastructure.

[SPEAKER_00]: Side.

[SPEAKER_00]: Your biggest design call was decoupling the mcp architecture, so multiple agente systems can share unified memory through a standalone process.

[SPEAKER_00]: Right, that's a mouthful, but you know what I'm saying.

[SPEAKER_00]: Rather than each one coupling to its own Python runtime, walk a student what problem that was solving and the key decision in the design that you made there.

[SPEAKER_01]: Around that time, when I was working on a cognizant working on this MCP piece, it's getting traction a lot.

[SPEAKER_01]: Even when I was just joining, it's like the cognizant has been a big open source project.

[SPEAKER_01]: With a lot of followers and the next big thing that was there was people were using this like graph to some as a personalized layer of memory, but the next big thing was shared memory across multiple people.

[SPEAKER_01]: Like you can think of like we already had some use cases where companies were using this as one layer that was shared like by all the employees and growing thing was like this mcp layer that people were using as an mcp was picking up and mcp was one of the most frequently used interfaces to interact with the memory layer

[SPEAKER_01]: and the bottleneck that I discovered and later on improved on was MCP was its own thing, stand alone.

[SPEAKER_01]: So you could have cogniz as a breath memory, or you could have it as a graph plus MCP, but there was not much of a decoupling there.

[SPEAKER_01]: So when you wanted to have one MCP memory for all the people to share, that was a kind of like a hard thing to make it work.

[SPEAKER_01]: So the next thing that I worked on was just decoupling it, making it possible for you to work with a company as one big shared memory, but like across multiple MCP clients.

[SPEAKER_01]: Yeah, I think I was I went a bit into like maybe the not so interesting technical jargon there, but like I think that was kind of like the way I got there.

[SPEAKER_00]: Well it's very interesting jargon.

[SPEAKER_00]: Yeah, don't short change yourself on that.

[SPEAKER_00]: That's interesting stuff.

[SPEAKER_00]: Okay, so that's actually diving a little bit further into the shared memory.

[SPEAKER_00]: Right, so you went through that process and I'm curious about like an example, right?

[SPEAKER_00]: Of an agent task that breaks when each agent has its own vector store.

[SPEAKER_00]: But works when there's this unified state that you're referring to through the decoupled MCP architecture, right?

[SPEAKER_00]: Well, and I'm curious, what's the actual mechanism that makes the difference there?

[SPEAKER_01]: really interesting use case that we came across, like with people that were using Cognito was multiple agentic systems that were built on different frameworks.

[SPEAKER_01]: You can think of like Google, agent, developer kit and line graph based agentic system that were actually like solving different problems, but really needed to share it understanding of the knowledge graph.

[SPEAKER_01]: And that's was the piece that I was breaking because up until that point, those were isolated from each other.

[SPEAKER_01]: But when you wanted to make those two systems work and build on shared knowledge graph, that's when you were actually blocked.

[SPEAKER_01]: And that's one of the first things that this thing that I worked on has unlocked.

[SPEAKER_01]: And even at that time, we were already supporting, I think,

[SPEAKER_01]: around five very popular genetic frameworks, and after that change that like I worked on, you could mix and match all of them and make them work concurrently on the same knowledge graph and make it like grow independently and concurrently.

[SPEAKER_00]: Super cool, that's really interesting.

[SPEAKER_00]: So in this space, most engineers that come from machine learning or an applications background, I'm primarily applications from my background.

[SPEAKER_00]: And you're coming at agent infrastructure from a peer distributed systems lens.

[SPEAKER_00]: What does that lens let you see that the ML native crowd is missing?

[SPEAKER_01]: I think in the Distribute Systems, first thing is in potency and, like, retries.

[SPEAKER_01]: So, across Distribute Systems, you may not always have things work.

[SPEAKER_01]: Things work very randomly.

[SPEAKER_01]: Back back then, I was also working in Cognit on this up.

[SPEAKER_01]: I'm up and coming platform that we were building.

[SPEAKER_01]: And the thing was, up until then, the memory layer was just on your local system.

[SPEAKER_01]: It was up there, you were the only person that was using it.

[SPEAKER_01]: Maybe multiple processes on your local machine that thou's pretty much it.

[SPEAKER_01]: But going forward, like with this platform and speeding up and like more and more use cases where people were like sharing the same memory layer across distributed computes, there were these pieces of distributed systems that were coming into the picture.

[SPEAKER_01]: network is not that reliable.

[SPEAKER_01]: The same things don't work from the first try.

[SPEAKER_01]: So you have to do retries.

[SPEAKER_01]: Then there's a competing request.

[SPEAKER_01]: So there come these pieces that become more and more important that we're from my in-for-a experience.

[SPEAKER_01]: So I would say like since like it impotency in retries and then a shared state across process, even on the same machine, that's where I think

[SPEAKER_01]: these tiny nuances that are a bit annoying when you're a developer, but I think those are the problems that you like, eventually would face when going into the distributive systems.

[SPEAKER_01]: I think that's kind of I think the pieces that are always grateful to have learned when I tried to work on carbon.

[SPEAKER_00]: Sure, now that makes sense, let me make sense how that transferred and gave you something of you that most people didn't have.

[SPEAKER_00]: Let's move into to beat.

[SPEAKER_00]: So the BDI, BDI is an injury and horror with SPAC startup and going after compliance audits, right?

[SPEAKER_00]: Which isn't the obvious first market for autonomous agents, but as a startup founder myself, it is something that's really an interesting space to go take on.

[SPEAKER_00]: You joined as a founding engineer in January and you've been shaping the technical court.

[SPEAKER_00]: From your seat,

[SPEAKER_00]: What makes compliance audits the right wedge for agent infrastructure?

[SPEAKER_00]: And I think it's a really interesting way to look at it.

[SPEAKER_00]: And what are the foundational decisions you're making today that will define what the product can do two years from now?

[SPEAKER_01]: I've been like an avid user of the autonomous systems.

[SPEAKER_01]: I was really keen on learning, or was it be going?

[SPEAKER_01]: And I was really like captivated by their mission when I was first applying them.

[SPEAKER_01]: from the first question, what makes compliance audit the right way for agent infrastructure is agency's days and autonomous systems in general.

[SPEAKER_01]: They're moving really fast.

[SPEAKER_01]: They're like with every new release, they're becoming more and more knowledgeable, much, much more smart than previous versions.

[SPEAKER_01]: And what I've learned first hand by building agent systems here at beat and in general is they are super good if you know how to leverage them.

[SPEAKER_01]: and so much so that I think I think I'm really confident that there's there are so many more spaces that agent systems will be able to work on part if not much better than humans going for it.

[SPEAKER_01]: And I think audits are like this already behind the wave, like meaning that the wave has become already surpassed the capabilities necessary to solve this very efficiently.

[SPEAKER_01]: And I think

[SPEAKER_01]: really good perks of being a founding engineer is you are really like a one of the foundational pieces of the team.

[SPEAKER_01]: So beyond being just a software engineer, I'm also taking part in the key decisions day-to-day.

[SPEAKER_01]: There are so many big important decisions that we make on like day-end and day-out.

[SPEAKER_01]: And I'm really grateful, and I think it's really full-filling work in that sense, because our decisions are actually literally guiding where this company is going, and feeling this kind of responsibility, and with some time later down the stream, seeing the results of your own decisions, it's really full-filling.

[SPEAKER_01]: So I think it's really a nice place to be both, like I think, as a company, but also as a person, as a founding engineer.

[SPEAKER_00]: Certainly, now that makes a lot of sense and your statement about it being behind the curve a little bit or that's not necessarily behind as a company, but the technology is there to solve that problem and solve it well, right?

[SPEAKER_00]: And so that's really cool to cool position to be in and kind of coming out as one of the front runners in solving compliance there.

[SPEAKER_00]: I'm curious, in that sort of space, in this sort of agentic era, you're obviously a professional and an expert in the space.

[SPEAKER_00]: Tell me, tell me a claim you have around agent infrastructure that most people in this space maybe would push back on.

[SPEAKER_00]: So what is your contrarian view maybe of agent infrastructure and tell me how you defended against a dissenting voice?

[SPEAKER_01]: That's a good one.

[SPEAKER_01]: I think I'm still maybe a dissenting voice in that sense.

[SPEAKER_01]: Maybe lately a little less, but still I think this is something that maybe some people would disagree with me, but maybe I'm also like this comes with a bit of a bias for you my background in infrastructure, but I firmly believe that, at least right now, Sandboxing is the only real safety boundary for autonomous agents.

[SPEAKER_01]: There's prompts, there's a lot of layers of safety that you can.

[SPEAKER_01]: leverage, but at the end of the day, especially with so many famous supply chain attacks that are happening these days, I firmly believe that Sandboxing is the only safety that you can actually rely on when things like when things don't go as planned.

[SPEAKER_01]: And I think maybe going forward this might change a bit, but right now I'm a firm believer that this might be right now

[SPEAKER_01]: This is also like some in a sense guides my decisions, especially like right now when working at beat because it guides guides my decisions with safety and mind and I might be a bit pedantic in that sense but you know you can never be too safe when it comes to especially I think client data and people like relying on you so I think that would be my claim that I would stick

[SPEAKER_00]: Yeah, I would say bye too.

[SPEAKER_00]: I'm right there with you.

[SPEAKER_00]: I think that makes sense.

[SPEAKER_00]: And you have to be careful, especially in compliance solution, right?

[SPEAKER_00]: People's data, people's processes, certifications, perhaps even audit results, things like that.

[SPEAKER_00]: That's really important.

[SPEAKER_00]: And you gotta be careful with all those things.

[SPEAKER_00]: So I'm certainly agree with you there.

[SPEAKER_00]: So with BDA and without breaking anything confidential, what's the hardest unsolved problem on your plate?

[SPEAKER_00]: Now we know what BDA is and how you're using agents to solve compliance problems.

[SPEAKER_00]: What's the biggest unsolved problem and how are you approaching it?

[SPEAKER_01]: One of the problems that I would probably flag, it's not something that I haven't solved per se, but I think...

[SPEAKER_01]: there's still room for growth.

[SPEAKER_01]: One of those things would for me would be Evel's.

[SPEAKER_01]: I I understand and I like work on it every day.

[SPEAKER_01]: I think there is always room for growth there because Evel's are the backbone of how you evaluate your systems performance and how you safeguard it against regressions and basically like some actual source of satity check for your

[SPEAKER_01]: I think the challenge there that I'm kind of like trying to solve, one of the things that Alma played out wanted to be better is there's this thing that you like the actual product that works with the sensitive data and everything and there's this e-viles which we try to make it like you try to make it as close as possible to the

[SPEAKER_01]: actual workflows and things that happen on production, but there's always this tiny gap because the data that agent systems are fed with, they're always unique each client, each use case, it has its own slide cork and you can try to replicate it locally, but there's still this

[SPEAKER_01]: Although I think we're in a good place, I'm always trying to chase that last 20% or maybe 1% but I would love to make it better.

[SPEAKER_01]: So I think that would be first thing that would come to my mind and I'm thinking about that.

[SPEAKER_00]: That makes sense.

[SPEAKER_00]: I make sense.

[SPEAKER_00]: Obviously it's full of problems too solve which is why the company exists and why you have the work set before you.

[SPEAKER_00]: But that's a good one.

[SPEAKER_00]: Okay, dot last question, two years from now, what's the piece of agent infrastructure that we'll consider obviously necessary, but doesn't exist yet.

[SPEAKER_00]: So like the electricity of the agent infrastructure if reach future or the water, right?

[SPEAKER_00]: What's going to be a utility who builds it and what does it look like?

[SPEAKER_01]: I think this is an interesting and maybe I'm kind of like stepping a bit a little too into the feature but I think going forward with more as you mentioned and as I like like touched on this wave, this curve is going to go further and with that mind there's going to be more more problems that agent systems will be able to solve efficiently.

[SPEAKER_01]: And with more more things being handled by agent systems, at some point there's going to be this necessity of regulating it a bit.

[SPEAKER_01]: But right now agent systems are like, there's their software engineers and software teams that are behind them.

[SPEAKER_01]: But going forward, I think there's going to be an SSD maybe.

[SPEAKER_01]: to have some sort of identity and capability registry for individual agents.

[SPEAKER_01]: So in that sense, you can think of individual agents being uniquely identifiable by their capabilities and also by the things that they've changed and worked on.

[SPEAKER_01]: So that you can think of maybe your opinion being like forrunner and does because if you think about it, I think with things being held by gender systems, you would probably at some point ask yourself, how do we find the responsible entity for the actions that this agenda gets to mass taken?

[SPEAKER_00]: And I think at some point there's going to be maybe discussion of, yeah, there's going to be needed for assigning identities to individual agents.

[SPEAKER_01]: And maybe I'm like, this is going a bit into like futuristic scenarios.

[SPEAKER_01]: But I think I can see this happening in the near future.

[SPEAKER_01]: So I think that's going to be probably one of the next steps in agentics systems infrastructure that's going to come into reality.

[SPEAKER_00]: That's interesting.

[SPEAKER_00]: Now I appreciate you sharing your view there and we'll pencil that down and see if it starts to come to life.

[SPEAKER_00]: And maybe we'll have you back on the show.

[SPEAKER_00]: So if it happens and we'll have a conversation about it.

[SPEAKER_00]: Thanks for being on the show today.

[SPEAKER_00]: You have had an interesting fruitful career spending time and reliability and then moving into more agentic space and not landing at BDI.

[SPEAKER_00]: It's clear that your experience is built on top of itself to set the stage for your success of this company and the problems you're solving.

[SPEAKER_00]: As well as while AI and the industry continues to mature from an agentic and model perspective, you're well positioned to keep your finger on the tabs, utilize the technology that comes out and solve problems for the compliance space.

[SPEAKER_00]: So, Dalit, again, I appreciate you being on the show today.

[SPEAKER_01]: Oh, thank you.

[SPEAKER_01]: Thanks for having me.

[SPEAKER_01]: It's been an interesting discussion.

[SPEAKER_01]: Hopefully, as mentioned, maybe if possible, like, into your style, we'll be able to come back to this, and ask us this in retrospect.

[SPEAKER_00]: Absolutely, discuss those predictions.

[SPEAKER_00]: We'll make sure and do that.

[SPEAKER_00]: You can clearly see how Dalits' career has led forward into a foundational moment for what he is building at BDI.

[SPEAKER_00]: The combination of reliability, infrastructure experience, and AI backbone, infrastructure don't let his well-position to solve the next big problem using the AI.

[SPEAKER_00]: Compliance.

[SPEAKER_00]: If you'd like to connect with Dalit, check the show notes for all the appropriate links.

[SPEAKER_00]: And thanks again for listening.