Triangle-Based Epistamai Uses AI to Remove Bias in Credit Decisions

Conversations about AI are everywhere, and often involve concerning topics: the threat of ChatGPT to academia, the ethical implications of AI in healthcare, or the impact of AI on the job market, for instance. Given these headline subjects, it’s easy to understand why some are uneasy about artificial intelligence.

But Epistamai’s CEO and Founder Chris Lam is working on algorithms and models that, hopefully, will improve people’s trust in AI systems. The models aim to reduce bias and discrimination in credit decisions and may eventually impact other industries as well. 

Lam poses the question, “How do we prevent a credit underwriting model from causing unlawful discrimination?” He adds, “I’m trying to figure out how AI systems become biased and unfair.” 

Lam founded Epistamai in 2022 and hopes to launch his product next year. Epistamai is a part of this year’s cohort at CED’s GRO Incubator, which is a 12-week program for early-stage startups. (We have also written about fellow cohort members Benevolist, Celestic, Pathstitch, Theralinq, TruPacta, and WikiELN.) 

Lam said GRO has been assisting him with customer discovery and finding his target market. It has also helped him refine his pitch and provided guidance for funding applications. So far, Epistamai is completely bootstrapped.

Epistamai’s model uses causal AI, which is a technique that seeks to leverage existing knowledge about how data is generated in order to model causal relationships in data—as opposed to modeling data based purely on statistical patterns and correlations. Through this approach, the goal is to use new sources of data to improve credit underwriting. 

The startup’s name comes from the Greek root “epistamai,” which means “to know or understand.” This speaks to Epistamai’s mission to determine the causes of discrimination and bias in credit decisions—a mission driven by Lam’s own background working at the Federal Reserve Bank of Chicago.

According to a 2023 study by Federal Reserve Bank of Kansas City economist Ying Lei Toh, traditional credit scores (like FICO) may not present the most accurate measure of a borrower’s creditworthiness. Toh argues that some minority groups are less likely to have their credit needs met through these traditional measures. Using alternative sources of data could help expand credit access for those underserved groups.

With Epistamai’s model, Lam aims to address both disparate treatment, which is intentional discrimination, and disparate impact, which is unintentional discrimination. Lam described disparate impact as a result of a neutral policy having an adverse impact on a certain group.

He said AI experts have been researching this problem for decades, trying to build an algorithm to address both disparate treatment and impact, but significant progress has not been made. To achieve his goal, Lam said he’s working on aligning his mathematical models with existing anti-discrimination laws.

“You want to be able to build AI systems that are compliant with the law, which is supposed to be a reflection of society’s values. So, that way you can build AI systems that are aligned to what society wants,” he said.

One of the reasons academia has not made much progress on causal AI in this sphere to date is due to politics, Lam said. Different political ideologies involve different interpretations of what is “fair,” so the AI must be able to model the root causes for why, say, Democrats and Republicans sometimes view the world in fundamentally different ways. 

So far, Lam has secured one patent on his algorithms and has three additional patents pending.