Jake Moilanen

I'm the CTO of a company called Iron Light.

We're a marketplace for tokenized assets to trade on, kind of like an exchange. But it's under a distinction called an ATS, Alternative Trading System. This thing about an exchange, but a lighter version. We're fully regulated. We have to deal with SEC and Fender all the time. So we have some interesting challenges when it comes to AI and how we actually use AI in a regulated type environment.

So in the regulated type environment, we're not talking about this, I'm going to trade you my board ape NFT for your fake copy of Dune. You're talking about actual securities.

That's right. So no cryptocurrencies or NFTs. These are bonds. These are real estate structure products,

mortgage backed securities. Everything you can imagine that's going to be that the financial industry trades. That's maybe not considered a typical stock.

So you're watching the private credit markets.

Private credit markets is a huge thing right now. We are very involved with that. It's a big, big deal for us. We solve a lot of liquidity issues that the private credit has with our system.

So what's keeping you up at night?

Well right now, actually the biggest thing that keeps me up right now is just what's going on in the UAE and Iran because we have a huge office over there. And I'm getting reports of every morning I get in and I hear about the bombings or some close call that one of my team members have had and trying to support the best I can, which is not exactly easy from over the side of the pond.

Yeah, when you're talking about a globalized economy, the fact that now data centers are legitimate military targets as a sea change.

We had to also rethink how we're doing our whole disaster recovery because losing a data center is a very real possibility out there right now. And so we've actually added a bunch of new deployment options. So if we need to go do disaster recovery, we can do it very quickly.

Tell me a little bit like you're dealing with a marketplace, a trading desk. So we're talking about an alternative to in layman's terms like a NASDAQ, who are the biggest customers at this?

It's actually all the same customers that would be on NASDAQ. So the same way that it's like a Morgan Stanley or a JPM would plug into NASDAQ and trade on their systems, they plug into ours the same way. And so if you're an end user on say your Charles Schwab account, you just see instead of typing Tesla and buying something Tesla, you type in one of our tickers and buy and sell like a normal, normal, looks like a normal stock. That's the goal.

Gotcha.

How is AI affecting your business?

Well one of the challenges we run into is because AI is not as deterministic and we're fully regulated, we have to be a little bit more thoughtful with how we even think about it. When you don't have luxury of having a bad deployment and be able to roll it back, say it's a CI CD and you have a bad deployment, we can't roll that back. We have a bad trade, we're liable for that trade. And sometimes these are could be for millions of dollars. And so that's a much different rigor we have to add into it.

So we still want to take advantage of it. And we've been really been tracking AI and been able to use it for our development as much as we can. And it's really back in about last fall, I finally said, I think the tools are good enough now that we can start using this.

And I think in December, we finally said, OK, let's go ahead and allow developers to do this with some guardrails. Because we have this regulated piece of software, we can't have that go and have AI just go wild on that. So I said, no actual code being written outside of the tab complete for those types of products. But for other products that are like our front end, well yeah, let's use that all day long. But what happened was in the course of within weeks, we saw a 4.5 X increase amount of code being developed per developer. That's net lines of code, which is a huge, huge number. But as you can imagine, that starts bringing up a lot of questions. OK, we have a lot more code now. That's a lot more code to review and make sure there's no problems. There's a lot more code that that could have a cybersecurity and lots more attack surface for that.

And how do you deal with all that? And I remember taking all this back to the team one day and I said, look, this is all the things are going on. Is this a problem?

And we started digging into it. And we started looking at what is actually being produced. And a lot of the stuff that we're starting to see was actually defensive coding. So a lot more error checking, a lot more unit tests, a lot more documentation.

And so that made me feel better. But still, this is a lot more code. There's a lot of problems that could be coming along with it.

And so we started to do more research on this. And one of the things we started realizing was that, OK, this looks like this is good. We're OK. But I still was not able to use it for our core product because we don't want AI doing work there. And our core product is a little bit different from, say,

a design pattern like MVC, model view controllers. It's one thing to have data being pulled out of a database, doing some type of business logic, and then going in and viewing it. It's pretty easy. It's a very verticalized thing for AI to go do. And it can do it pretty well. But our systems process too many messages per second at 20 microseconds per transaction. So the same speed as NASDAQ.

There's a lot of interaction that happens in between going kind of horizontally.

And this engineering is where AI starts failing. It's not quite there yet. But we still want to take advantage of it. So one of the guys on my team, his name's Ryan Charles, he had this great insight. He realized that AI was not good at-- we can't always use AI just for engineering. But really, it's to improve how we think about things and improve the speed of thinking. And it's really more for the research.

And he developed this technique that he coined, Research-Driven Development, RDD. And the idea is this loop where it goes-- we go and define a goal of what we want to achieve. Not a plan, a goal. And then we come up with experiments, not tasks for AI to go do. And then we go and log everything. And we go log all the thought processes. We have the AI explain exactly what the thought process was. And we log it in an immutable form. So we have that record forever.

This is also doubly important for us. We fall under something called RegSCI, which just means it's this very high bar that exchanges have to be built towards to prove why we did everything we did within the software. So now with this immutable log of everything that occurs. But now it has the extra benefit that we have the thought process for all future agents to go read and review, as well as our team and regulators go review for how we got to a certain conclusion. And we've started using this in a lot of different fashions. And we've used it for speeding up our matching engine to figure out what is a better database migration tool. And we've been able to really iterate and get much quicker and take advantage of AI in this one part of our code base that we couldn't do it normally. So it's been a huge benefit from that respect.

It almost sounds like you're applying the scientific method.

It's precisely it. It's a scientific method.

That's fascinating because the SDLC, you look at Scrum, 30 year old tech, what the framework you're using to do this. You discover very quickly that productivity is not the same as outcome.

That's right.

And so once you start seeing that you were measuring the right thing, but the right thing isn't the actual right thing eventually that you get to. And so now that you're doing outcome based development, is that really a sea change from where the whole SDLC was before?

It was. We're actually going through the process of rethinking our whole SDLC and how does it need to be adjusted for this new type of techniques. And we've made some adjustments to our SDLC. We've looked at, we have always have an AI code reviewer, which is great. And we have it go in and put in comments directly into our pull requests.

So they actually have a log of all the questions they have. And it comes up with lots of great questions a lot of times. Why do we do something? And it's been very, very helpful. We've also made a change that not all code reviews are equal.

If it's a core piece of code that we have to have multiple piece of eyes on it still. And then other parts of the code, like say front end software, doesn't matter. It doesn't need as much rigor when it comes to the actual code review. And so we've kind of teared out our different ways of thinking about code reviews since we have so much more code being committed. We need the humans, our leads, to be focused on the ones that really matter. And once we made that adjustment and said that not all code was equal in code reviews it's been a fundamental change is how far we can actually how fast we get through things. And we haven't seen any dip in quality whatsoever because of it.

Right. Prioritization has always been part of the software development. So it just seems like the code review portion of it is where the bottleneck is. And you've just gotten better at triaging.

That's right.

Interesting. How big is your team?

We're small. We have the whole company is 27 people. We have eight engineers.

And how long have you been around?

2023.

Okay. So these tools to your point, six, eight month, that's really where you're starting to see the acceleration there. How have these new tools affected the way that you consider like,

let's say, and I'm just putting that business is going great. You want to grow. Does growing include hiring new people or simply doubling down on these tools? Do you need more people?

It's a combination of both. We have a team that's dedicated just to AI, building tools for our teams to make them more effective, faster. And we've been able to not have some hires because of that. But on engineering, we have not seen any, if anything, I need to hire way more engineers.

That's one area where I've not seen us pulling back on hiring because of it. Now what's interesting is that our hiring practices have changed quite a bit though.

We used to kind of look at just like a raw talent. We actually have now focused on people that are more product focused. And what I'm finding is people that have a background on product do dramatically better. One of my dear friends, he runs a 250 person development org. And he's finding that the people that are the best now are actually the product people on the product team are producing better results than the actual developers. A product person doesn't have to go in and try translating and put into a specs what they're thinking. They can just have the model go and do it for them. And then they can keep iterating very quickly on that to get to the result that they're looking for. And by having that product mindset, you can get right to what actually matters faster than you can get a typical non-product focused engineer.

Is it easier for a product guy to learn the tech or is it easier for the tech guy, the coder to learn product?

Definitely product guy learning the tech now, which for a long time it was flipped.

So that changes your hiring approach. Do you prioritize product? It feels like more of an in-person versus a like do you prioritize in-person remote? How does that work?

Yeah, we're all in office. We I mean, hiring has been a real problem. The amount of people using AI during the interview process on the first Zoom calls has been incredible. By my measure, it's been about 20 percent are actually trying to use an AI tool like Cluely, for instance, to try getting around, trying to do better in the interview. And we've had to develop new techniques to identify people trying to use these things. And one of the easiest ways of doing it is when you have them go through some detail and some question and then you go back about five minutes later and ask them to reiterate what they were talking about, go into more detail about some specific thing. But you're very vague about what you're talking about that they know what you're talking about. But the Cluely or some of these other AI tools don't. And then you'll find out pretty quickly if they actually are using a tool or not.

I don't trust anybody on anything online until I met the person.

Everything has everyone has to be in person at least once for us to to ever make an offer.

Yeah, that makes sense. And you're in a regulated industry. So you know, my assumption is that you have to be like you can't hire overseas. You have to be local.

We hire an UAE for the Emirates for for developers there. But outside of that, everyone is here in Austin.

How big is the pooling UAE at this point?

It's fluctuating quite dramatically right now. Yeah. But actually there's a lot of there's actually a lot of people there. It's mostly the most of expats that are over there right now. And and there actually is a pretty decent pool. But it's yeah, it's been I'm hearing from a lot of people wanting to leave right now, which is understandable.

Yeah. I mean, the Israelis have more coders, but also the Israelis are spending out their time in bunkers right now. Yeah.

How let's let's get into what are you the most excited about right now? With all of this new tech with I mean, you're deep in financial markets, which is causing a lot of other people anxiety, like with these new tools, with the way things are going, anything that you're really excited, you're looking forward to.

Yeah. I mean, one of the things that I think I'm most excited about is that we can add in. It's not the core development portion of our of our company. It's the ability to add on tools that improve. It's very customized for the larger part of the company. So, for instance, we we have a tool now that will go before and before any time we have a have a meeting if it's an external meeting, we'll go through and look to see who the participants are and then go pull a quick bio on them and also figure out what try making a determination of what they want to talk about, what was successful like in the meeting for them and what we should bring to the to the to the actual meeting itself. And that's been just a simple tool like that. So just a little Google calendar extension has been invaluable and we've we've have seen things in a different different light.

We have things that we we use even for, you know, for how we actually think about doing our slack interactions. And we have tools that we built to see where there might be some conflict between team members and that we're starting an idea of what what is where there might be problems from a management perspective we need to go focus on. And there's all

these tracking fights in Slack and

we are so all a sentiment analysis based off of these are all public channels. Whenever there is a thread between two people or more more than that, we will go and track to see what is the how is their interaction. Are they fighting or is this is this cohesive? There's always a little bit of friction. But, you know, this we've actually been able to go head off a number of different problems within the overall company because of this. And these types of tools just wouldn't have the time to go do in a normal in a normal course of business. But with with A.I. And we can go develop these things very quickly. They're not perfect. That's fine. And it's I mean, A.I. is really is a lot of these tools are fast fashion. They're great, but they're not scalable. But for this type of stuff, these like small tools, they're perfect. We've also we've gone back to SCLC real quick. We've also changed how we even think about some of our architecture. And a lot of places we start looking at doing a microservice architecture. So we can do things in smaller chunks that that are easier for A.I. to go process and go figure out versus something that's much more complex and monolithic. So that there's been lots of interesting ways that this is really kind of allowed us to to be be better as a company.

What, if anything, is keeping you up at night?

Well, I mean, from obviously the the the war is going on right now. The outset of that, you know, I'm always concerned that somehow we have some something that gets slipped into our into our code base that that A.I. produces that's wrong and we slip through the cracks.

The Superman three problem.

Yes, pretty much. And which is always always a concern.

The least reference Superman would be. Do people. Richard Pryor, that's my favorite. I don't know if people reference that Superman three problem or they reference the office space reference.

I think I hear a bit of both, but yeah.

Now, for those people, the financial industry.

Is I feel like in banking, fintech, all of that stuff are leveraging these tools, but they're having to be more methodical simply because of the regulation around.

That's right.

What kind of changes do you see happening in that space that people aren't aware of? But they're going to be.

Yeah, I mean, we're seeing this across every single bank I'm talking to has A.I. team and they are all figuring out how to improve improve their company by by using these tools. As you mentioned, one of the challenges that they run into is that it's they have to be a little more thoughtful when it comes to how they actually use it and having having we always tell them to keep this record of like exactly all the thought processes that the A.I. tools are doing so you can go back and reference it and making sure that regulators can understand how you got to that point because it's not you can't go back and ask A.I. later, which is unlike a human, if a regular can go back and say, well, why did you do this this thing? It's not as easy for A.I. to go do in the moment. So that's one thing we've been we've been seeing. And I think it's one thing that I see a lot of doing at this point.

What we're trying to do right now is I'm trying to do the front end research side and talking to people. And then we hired a full time academic machine learning PhD to basically track every change, every research paper, everything that's coming out constantly just to continuously build out new models.

And we have agents that are running to go as regulation comes out on all these different jurisdictions. It's difficult to track all of it, especially how quickly the regulations are changing in our space. And so we have agents that are constantly just looking for that and go and flag it. It was funny. We had this we had this there's this SEC document came out, FAQ document came out that we wrote a little bit of part of it. And we had no idea until it came out. But when our agents detected it and said, oh, by the way, this looks like it looks like something is applicable to us. And like, oh, well, actually, we actually wrote that. That's great. You know, so it's like these tools are effective for us to have a bit better breadth of what's actually going on in the world and keeping and keeping tabs on it.

Are you purely national regulation or do you have to give a state level?

We're largely national and federal. We have every jurisdiction is slightly different. And and UAE, it's different. They have different regulating bodies depending on what we're working on that we have to deal with.