Sardine, an agentic risk platform focused on combating fraud and financial crime, has set up Sardine AI Labs, an applied research unit that will build foundation models trained on financial behaviour, backed by $375,000 in research fellowships.
The new group will investigate how AI can interpret patterns in financial activity, predict emerging attack methods and support risk teams in reaching sounder decisions. Sardine is also inviting applications for as many as five fellowships, which will go to independent researchers who can build on the lab’s initial findings.
The company argues that the coming generation of foundation model breakthroughs will emerge from specialist AI neolabs that hold proprietary, real-world data. Fraud prevention illustrates the point, as even prominent academic researchers frequently cannot obtain datasets at the scale required to move the discipline forward.
Sardine treats payment and account access activity as a language with its own grammar, in which transaction records, devices and IP addresses combine into behavioural profiles that models can examine for anomalies and signs of fraud.
Financial crime prevention continues to weigh on banks and other institutions, pushing up processing costs across card, ACH and wire payments. Sardine highlights sanctions screening as one pressure point: customers whose names or identities resemble entries on sanctions lists can be wrongly flagged, stalling onboarding for weeks during manual checks.
Establishing whether two records belong to the same person remains difficult. Spotting money laundering, terrorist financing and drug trafficking payments within billions of transactions poses a second challenge, as current tools can block legitimate payments while still missing illicit ones, leaving institutions open to regulatory and financial penalties.
The lab’s research agenda covers modelling very long sequences of transaction and user behaviour events, transfer learning, adversarial robustness and explainability. Drawing on Sardine’s device, identity, behavioural and transaction data, it aims to build models capable of recognising attacks they were never specifically trained to detect.
Sardine AI Labs has also released early results from a transformer-based card model trained on roughly one billion transactions gathered over two years from more than a dozen card issuers.
The model targets the cold-start problem, where newly launched issuers have too little data to catch fraud reliably yet attract early attention from fraud rings. When tested on issuers excluded entirely from its training data, the model lifted fraud detection accuracy by 68% for a consumer card issuer and 41% for a business card issuer against standard machine-learning methods. Improvements held across consumer, business and global cross-border issuers, indicating the model can generalise across varied card programmes.
Sardine CEO and co-founder Soups Ranjan said, “We are uniquely positioned to train a model purpose-built for risk because we sit on the industry’s fastest growing fraud and fincrime dataset, which spans more than 6.5B devices, 441M consumers, 3.4M businesses, 6.6 billion transactions and $1.8 trillion in payments. Sardine AI Labs will focus on production-grade models that meet real world latency, governance, and explainability requirements.”
Sardine head of data science Niranjan Shetty said, “The most important finding is that foundation models can learn directly from a user’s transaction behavior, allowing us to identify fraud more accurately than approaches that reduce that behavior to a set of tabular features. Because these patterns transfer across financial institutions, the model is not limited to a single card program. The next step is to make this intelligence fast, explainable, and reliable enough to support real-world risk decisions.”
Ranjan added, “The frontier AI labs have shown the immense benefits of having multi-modal AI models which allow you to extract intelligence across different modalities of audio, video and text.
“We think that the next unlock in fraud and fincrime prevention would occur from creating a multi-modal AI that goes across device, identity, behavioral, and transaction data. This is what excites me the most about the potential ahead with our AI Lab and I can’t wait to see what our research fellows build with us.”
Copyright © 2026 RegTech Analyst
