Alex Shieh is using AI to hunt $600B in government fraud — and gets paid only if he wins

and then filed an insurance claim, which was then disproven by, I believe, some sort of uh biologist who knew about bears and what they look like on camera and was able to debunk it. We were able to debunk it, too, because it looked ridiculous. Anyway, we have Alex from the Anti-fraud Company back on TVPN here for the Teal Fellowship Giga Stream. Welcome to the show, Alex. How are you doing?

Great to see you. Suited up.

Suited up. Fantastic.

Looking sharp.

I cannot look like a fraudster when I'm catching fraudsters.

Have you been dailying a suit?

Uh, not not every day. Only only when we're doing formal things. Yes. Yes, indeed.

So, uh, you've been on the show before, but sort of reintroduce the company and the shape of the business and how you think this will play out. Then I want sort of the update on where things are, what traction looks like, how you're actually applying the technology, and then I'm sure there'll be a ton of follow-on questions.

Yeah. So, it's it's very simple. We use we're building AI models to detect fraud when it happens, and then we go and sue the fraudsters, and we only make money on contingency when we actually get a recovery for the government. We're able to keep 15 to 30% under these whistleblower laws.

Interesting. Yeah. And uh have you received a whistleblower uh bounty yet or is that something that's a goal for the next year or two?

That that is a goal. Maybe not for the next year or two because the US legal system is is quite slow. So it might take three to four years before we see our first dollar of revenue actually.

Okay. So

how much uh how much can you pull forward? I know that the plan is to use AI, but how much can you pull forward by just rolling up your sleeves and doing it yourself or hiring humans to sort of follow the same process? Is this a problem that can only be attacked with AI like a recommendation feed? Like you could never curate like the Tik Tok feed for every individual on that platform without AI, but you can certainly, you know, go and take a photo instead of using AI. uh how how how dependent is this on on AI uh from the very first project

right so I would say that people have always been able to find fraud manually but usually either they luck into it or they had a tip for from somebody on the inside and we're sort of trying to eliminate that bottleneck essentially we want to find the fraud using AI because we we're we're essentially building the palunteer of fraud detection we're we're tapping into all these sources of data and once Once we have a good idea of what's going on, um, then we'll send out human investigators again, go talk to sources, but we want to be going outbound. We don't want to have to wait for people to come to us because that's just that's just less less efficient. When there's 600 billion dollars in fraud, you're not going to get tips for most of it. And if you want to solve it, then this is sort of the way you have to go.

I remember during the uh the COVID stimulus checks, there was someone who got a a check for a company. What what were those called the stimulus checks? It was it was specifically for business loans,

SBA

SBA loans, something like that. But those those loans, one of them was just called like Ford Raptor LLC,

and then someone dug into it and lo and behold, like someone set up a fake

fake LLC just to buy a Ford Raptor. Uh, and that one seemed extremely obvious. Uh, but I'm wondering how much is available in terms of public records versus stuff that you plan on doing FOYA requests for to get new sources to comb over.

That's a great question. So, we we use So, this sort of stems off of like the the OSENT community, open source intelligent.

Sure.

Which is where people go around and they look at publicly available data and and when you're doing pure OSENT, you're only using uh again open- source stuff that's publicly available. That limits you. So we do get closed sources too, but I guess we we want to start there. We view it like a funnel. So starting with the the open- source information lets us cast the widest net. Um, of course, some of it we're going to need to do do other things like you mentioned. Foye is one way that you can get uh private information, but it's it's really like you have to know again what what you're requesting from the government or what you what sort of private data source you want to buy or what human you want to go out and talk to. And so to to get the highle view uh you you do need to be essentially just ingesting a bunch of stuff. So one one analogy we like to make is is the story of the the blind men and the elephant, right? Is there's like five five blind men and they're each touching a different part of the elephant. The guy who's touching the legs thinks it's a tree. The guy who's touching the trunk thinks it's a rope, right? Is fraud is kind of like this. It always leaves uh traces, but it's it's somewhat difficult to put them back together. And that's why we just want to have as many different essentially like sensors on on the government databases on contracts on all these data sets. And that gives us a sort of a a good look into what's going on. Then we then we ontologize it, run it through models, make make graphs and see if it matches on to any known patterns of fraud that we're searching for.

How many how many like projects are you guys taking on at any one point? Because I'm sure every single day there's like new ideas, new opportunities. Sounds like a very very big market unfortunately

quite quite a few projects at a at at a time. Um and and again we're just looking everywhere. Um I mean like you mentioned small business administration is is a great place to look with all these loans, defense um just anywhere the government spends money. Um there there is undoubtedly going to be people who are doing it unethical way. Oh absolutely healthcare. Yes. I I mean like we spend trillions on healthcare. We spend trillions on on defense. Um there there's a lot of stuff going on in these places.

And what's going on in Sacramento right now? Is there some new legislation that's trying to make it harder to do journalism around uh fraud? Did I see that?

I believe so. I think this was a response to Nick Shirley, I believe.

Oh, really?

I think it's called the Nick Shirley.

Wow. Okay. Right.

I don't know if it's actually called that, but that's what people on the internet are calling it

over the over the break over the holiday break.

Yes. No, that that that that's my understanding. Obviously, Nick first did like the the big expose about the the the the daycarees in Minnesota which were getting funded by Medicaid or something of that sort. And so that that also helped drive a lot of awareness um and interest in our company as well because I mean it's just sort of visceral. Um, I guess first there was Doge where people just got really mad about how their money was being spent and then the the the daycare sort of was the was the was the next step. And I think there's just a big anti-fraud moment um in the United States these days where people are are just upset about the their dollars aren't going far enough, right? I mean like at 600 bill uh billion dollars a year. That's like 7 8% of the federal budget. So that that means like a good chunk of your hard-earned tax dollar just being stolen by people who are um contracting with the government in the wrong way. And I mean we need to restore trust in our institutions. We need to have good governance have have people have faith in our institutions. And I think that um again aligning the incentives having sort of private actors come in and do this and rewarding them when they're able to claw back money for taxpayers. The most of most of the money that does claw back will go back to the the Treasury and to the taxpayers. I think this this is kind of a no-brainer.

How do you think about uh the actual application of AI tools and models uh in this particular case? I understand that you're you know sort of creating a database or mirror uh a data lake of all the different sources. Uh but then do you need a bunch of examples and to fine-tune a model? Do you just need to load examples of red flags into the context window? How far down are you on like the AI research side of understanding the problem? Because I I can imagine like with a frontier model you can go and if you give it a lot of examples and and data points you could potentially just one shot like a detection and then it's just applying it at scale is the problem. But how are you thinking about applying the actual technology?

So there are essentially two steps here though. The first step is that a lot of this data is unstructured. A lot of this like is is documents and so LLMs really speed that up. It's sort of uh getting structured entities out of uh the unstructured data and then we we ontologize this. We have like this company has this contract has this relationship with this employee etc. We build sort of this this ontology knowledge graph like thing and then we have the second layer the the rules model. Essentially my co-founders are are both lawyers have a legal background. They've been working on sort of the fraud issue in in the legal sphere. And so they know what that looks like. And with their input and with their expertise, we're able to essentially develop this rules model, this rules layer, compare that to what we're seeing in the real world and flag things that look like violations in real time or at least stuff that could be violations and we need to get more information on.

Did you drop out? Because if you have two lawyer co-founders, did they drop out? Like what's the what's the story?

Who's the dropout?

Who's the dropout?

I'm the dropout. my my co-founders have have JDs unfortunately and they're also a bit older so they're not eligible for the teal fellowship unfortunately but um

I did I did drop out of Brown um uh in in 2025 to to work at Palunteer and then start this company.

That's great. That's great. Uh well, good luck and thank you for everything that you do.

How big is the team now?

Uh we have 11 full-time employees, but we're hiring engineers. If you're an AI engineer that wants to work on this uh problem, visit anti-fraudcomp.com and and we'd love to work with you.

It's a great name.

Uh well, good luck out there. The good fight and uh let us know when you catch a big culprit so you can come on and tell the story because I'm sure it'll be riveting.

Yes, indeed. Thanks, fellas.

We'll talk to you soon. Goodbye.

It would be funny if he was investigating a company to have the founder on. Oh,

we're interviewing them and then Alex joins the call.

Sort of a Yeah, the catch a predator.

That's what you want to do with fraud. This is a new media thing. This is the new media opportunity.

New media for for pump and dumps and schemes and all sorts of stuff and rug rug pulls uh that inadvertently happened during the NFT and and crypto boom because there were a number of founders who went on shows and then it was revealed. I mean, the the famous one is Joe Weisenthal and uh and uh and some Bloomberg reporters talking to SBF, I believe, on OddLots, and they asked him like, "So, this is like a black box that you put money in, you just get more money out." And uh he was like, "Yeah, exactly. Exactly." He was like,

SPF was on OddLots?

I'm pretty I'm pretty sure it was OddLots uh with uh Yeah, I I know I know Joe was on uh that that podcast with SPF. Um and uh yeah

and and and he basically describes a Ponzi scheme.

This is crazy. SPF and Matt Lavine Matt Lavine on

and Matt Lavine asks like so he describes a Ponzi scheme and SPF basically just says like exactly like that's why it's good. It's like you you put more money in then you and then it grows and then it's this magical system. He was like describing like crazy DeFi schemes. Uh it was a it was a rough time. It was a crazy time. Uh but we