Sardine launches AI research lab and $375,000 fellowships to push fraud detection with foundation models

Sardine has set up Sardine AI Labs, a dedicated applied research unit, and attached $375,000 in research fellowships to it. The point is straightforward: use foundation models to read payment behavior the way fraud rings already do, and give risk teams something faster than rule sets and tired heuristics.

1. Sardine AI Labs will focus on how AI models can interpret financial behavior as a form of language, predict emerging attack methods and help risk teams make better decisions. The research agenda includes modelling extremely long sequences of transaction and user events, transfer learning, adversarial robustness and explainability.
2. As part of the launch, Sardine is inviting applications for up to five fellowships for independent researchers. The fellowships are meant to extend the lab’s initial findings, which is a neat way of saying the company wants outside brains working on a problem that has not been solved by throwing more rules at it.
3. Sardine says the next breakthroughs in foundation models will come from specialist AI “neolabs” with proprietary, real-world data. In fraud prevention, that matters because even prominent academics often cannot get datasets large enough to move the field forward.
4. The company’s thesis is that payment and account access activity already has recognizable patterns: transaction records, devices and IP addresses can be combined into behavioral profiles that foundation models examine for irregularities. That same logic applies to sanctions screening, account opening and suspicious-payment detection, where false positives still slow onboarding and manual review.
5. Sardine AI Labs has already published early results from a transformer-based foundation model trained on roughly one billion transactions gathered over two years from more than a dozen card issuers. The model is designed to solve the cold-start problem for newly launched issuers, which are often targeted by fraud rings probing fresh products before the controls have enough history.
6. In testing on issuers left entirely out of the training data, the model improved fraud detection accuracy by 68% for a consumer card issuer and 41% for a business card issuer versus conventional machine learning methods. Sardine says those gains held across consumer, business and global cross-border issuers, which is the part that will matter most to PSPs and issuers wondering whether the model only works in a lab with friendly data.

For high-risk merchants and their payment providers, the practical message is clear: the cost of fraud prevention is still rising across card, ACH and wire payments, and the bottleneck is increasingly data quality plus model adaptability, not just more manual review headcount.