Transforming capital markets with AI

Former Meta data scientist and Citadel trader Richin Kabra recently joined our stealth fintech venture as CTO.

Richin Kabra is one of the newest founders to join the Merantix Capital portfolio, having recently teamed up with our founder-in-residence Anouk Moll on a new London-based venture at the intersection of AI and capital markets. 

Their fintech company is currently in stealth mode — but we couldn’t be more excited about what they’re working on. And we couldn’t be more excited about Richin, who joined as co-founder and CTO this summer after working as a senior data scientist at Meta for advertising auction science and previously as a senior quantitative trader at Citadel Securities. 

I recently sat down with Richin for a chat about his learnings from his time on Wall Street, what AI brings to modern finance, and the challenges of building in fintech right now. Our edited conversation is below. 

You began your career on Wall Street, eventually working on the ETF business at Citadel Securities, the American market making giant. What did you see there and did it make you think about becoming a fintech entrepreneur yourself?

Almost 20% of all U.S. equity trades are executed through Citadel Securities, yet they only had a small presence in ETFs.

It was almost like starting a new venture within a well-established giant. While I was a quant trader running the emerging market ETF desk, I also took on roles as a business development agent and an exchange connectivity agent. I was in discussions with the Johannesburg Stock Exchange to build connectivity there, and I was also reaching out to the Indonesian Stock Exchange, where connectivity hadn’t yet been established. We had to consider the amount of risk capital we wanted to deploy in each of these exchanges, essentially thinking about product-market fit.

This approach translates well to entrepreneurship: you have a roadmap, a sense of product fit, and you’re continuously testing adjustments in capital deployment to maximize returns and gains. That made me think - if I can do this for a behemoth like Citadel Securities, I could do it for something smaller as well, something I could call my own and engage with more deeply as a product.

What opportunities did you see at the intersection of AI and finance? 

There is a backlog of data processes and harvesting that is done very manually, including data entry and numerous mid-office risk compliance factors. All of these have a structural element to them. Anytime there is an input with structured data fields, I believe AI can handle that much faster than human intelligence, simply because it’s a repetitive activity involving structured data. Improved operational processes are definitely one pillar.

The second pillar, I would say, is risk modeling. We've repeatedly seen instances where correlations in financial assets have been mis-modeled. This was a major issue in the 2008 crisis, where the U.S. mortgage crisis stemmed from a miscalculation in the correlation across mortgages.

Do you think AI could have helped avoid that financial crisis?

I think improved models can surpass human intelligence. In other words, AI can bring an incremental improvement to risk modeling, which currently relies on two-dimensional models with standard deviation and time as the two axes, merely predicting changes in standard deviation.

We’re not just talking about changes in risk. We’re asking: if this were the scenario, what would the other influences be? We’re considering factors like housing credit, employment credit, and other opportunities in the landscape. How would that impact employment opportunities? How would it affect inflation? How would the Fed adjust interest rates in response?

Think of it as a game of chess. Traditional models take a single step - if a player makes a move, what would you do next? AI, however, can go much further, predicting up to 13 steps ahead to anticipate moves and counter-moves in response to an opponent.

After Wall Street you moved to Meta working on advertising. That seems like a pretty big jump.

If you think of ads as an asset class, they’re quite similar to traditional asset classes. You have a buyer, a seller, and an exchange. In traditional asset classes, the buyer might be someone in the retail world who wants to buy Apple stock. The seller is typically a market maker who’s willing to sell that Apple stock, and the exchange is usually the London Stock Exchange or the New York Stock Exchange.

The ads market works in much the same way. The buyer is the advertiser, and the seller is your screen time. The exchange is typically run by ad networks like Google or Meta.

What’s particularly interesting about Meta is that when you control the pricing for the auction system that connects buyers and sellers in an ad auction, you're heavily involved in regulation to ensure pricing is clear and justifiable to users, explaining why they’re seeing certain ads. Additionally, you aim to make the system as efficient as possible in terms of targeting.

We used user data to predict the best interactions for users — what they wanted to see. And for advertisers — what they wanted to show users. It’s a double optimization process. We would predict the best return for advertisers and the best return for users, then combine these predictions in a weighted mechanism to show users the most relevant ads. This way, users don’t dislike the platform, and advertisers get the best value for their investment.

What are the biggest technical challenges right now within building in fintech and AI? 

I can probably answer that in three pillars.

The first would be data protection and privacy. If you’re looking into consumer credit lending, you’ll want to consider data privacy, especially when it comes to protected classes—housing, employment, and credit—and avoid using those protected classes. I also believe gender should be considered a protected class. It's crucial not to use that data to build proprietary models for individual clients.

The second pillar would be explaining models. When you run AI in the background, everything often becomes a black-box structure. The outcomes are very different from the inputs because there are significant operations happening in between.

The factors that contribute to these operations aren’t always explainable because there aren’t always human-understandable factors. For example, when assigning a loan or credit to someone for purchasing a house, in traditional models, you’d look at employment, wages, and any other mortgages held by the person. But in an AI model, these factors are weighted together in a complex way.

This leads to a convoluted mesh where it becomes difficult to describe what’s actually being used to an operator, regulator, or compliance officer.

The third piece I’d touch on is initiative change. FinTech has been around for a much shorter time than traditional finance or tech industries. When it comes to adjusting to new models and transitioning to different infrastructure, you need to upskill people. You also need to convince regulators that these new models are solid. When risks and balances aren’t in place, failures can occur, and regulators often focus on these single failure events as points to revisit, which can slow progress. 

How have you found working with Merantix Capital and being part of the larger Merantix community?

I just came out of a coffee chat on model implementation, and everyone was super friendly and welcoming to a new person. It’s been incredible. Having a campus environment feels very wholesome. It brings you back to a time when you were in college with everything around you. Accessibility really boosts productivity.