Financial crime fraudsters constantly adapt their methods and move across payment channels. Sardine is responding with a new research initiative focused on using artificial intelligence (AI) to identify these patterns earlier.
On September 30, the financial crime prevention company launched Sardine AI Labs, an applied research group dedicated to developing AI models for fraud and financial crime prevention. The company said the lab will focus on models that can work in real-world financial systems, where speed, explainability and governance are important.
Key points
- 68% improvement in fraud detection accuracy for a consumer card issuer in early testing.
- 41% improvement for a business card issuer.
- The first model was trained on about 1 billion transactions from more than a dozen card issuers.
- Sardine is offering $375,000 in research fellowships to up to five independent researchers.
- The research will explore how AI can detect financial attacks it has not previously been trained to identify.
One of the first areas being examined is the so-called cold-start problem faced by new card issuers. When a new card programme launches, it has limited historical data to train fraud detection systems. That can make it difficult to distinguish legitimate transactions from fraudulent activity, while new financial products can also attract fraud rings.
Sardine said its first foundation model was trained on transaction sequences and then used to provide additional features to an existing fraud detection system.
“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.”
Niranjan Shetty, Head of Data Science at Sardine
In testing, the model was evaluated on card issuers that had been excluded from its training data. Sardine reported a 68% improvement in fraud detection accuracy for a consumer card issuer and a 41% improvement for a business card issuer. The company also reported that the model helped detect 35% more fraud using the same number of transactions. One must note, though, that these are Sardine’s early results, and they are self-reported, rather than an independently verified industry benchmark.
The new lab plans to study several areas, including sequential modelling, transfer learning, adversarial robustness and explainability. The broader goal is to combine different types of signals, including device, identity, behavioural and transaction data, to understand financial activity as a connected sequence rather than looking at individual transactions in isolation.
This could be particularly relevant to money laundering and other forms of financial crime, where suspicious activity can be spread across multiple transactions, accounts or institutions. Sardine says its research will also examine whether models can generalise across different financial institutions and operate with the low latency required for real-time risk decisions.
$375,000 research fellowship programme
Sardine is also opening its research programme to academics. The company plans to award up to five fellowships, with a total of $375,000 available to independent researchers at universities in the US and Canada.
The research topics include improving AI models across different card issuers, working with multiple streams of financial activity, identity matching and detecting money-laundering patterns.
Applications are open until December 13, 2026, with the first group of fellows expected to be announced in January 2027.
