What's Happening?
TypeSafe AI's new System One model, Jev, has shown significant success in real-time fraud detection, particularly in identifying fraud rings. In tests conducted by Sardine, a company assisting banks and fintechs with legitimacy decisions, Jev caught 27
out of 29 fraudulent transactions, achieving a 93% catch rate. This performance surpassed a rules-based approach (90% catch rate) and a general-purpose Large Language Model (LLM), Claude Sonnet (62% catch rate). Jev operates at approximately 1/25th the cost and with a response time of around 100 milliseconds per call, making it suitable for high-volume, real-time payment environments where traditional LLMs are too slow and expensive. The model's consistency in judging transactions, even with varying card usage within a fraud ring, was a key factor in its higher accuracy. Jev is designed to be called by software, providing quick, consistent judgments on structured evidence.
Why It's Important?
The financial sector, including banks, fintechs, and payment companies, faces an escalating challenge from sophisticated fraud tactics. The ability to detect and prevent fraud in real-time is critical, especially with the increasing speed of digital transactions. Jev's high accuracy and efficiency offer a significant advancement in protecting financial operations and customer accounts. Its low cost and rapid processing time make it a viable solution for institutions that need to make thousands of decisions per minute without compromising customer experience. By effectively identifying fraud rings and suspicious activities like name/email mismatches, Jev can help reduce financial losses, enhance security, and maintain trust in digital payment systems. This technology is particularly important for preventing card-not-present fraud and ensuring the integrity of new business accounts.
What's Next?
Sardine plans to further refine Jev's capabilities by decomposing broad questions into narrower ones, such as assessing bursts of activity or linked accounts, to potentially sharpen fraud scores even more. The model is also being integrated into Sardine Flow to automate Know Your Business (KYB) risk checks, turning findings into structured answers for customer agents. This integration aims to streamline compliance processes and provide actionable insights for other software systems. Furthermore, Jev's consistent judgment is expected to improve the quality of labeled cases, which in turn will lead to stronger rule recommendations for fraud detection. The ongoing development and testing of Jev in various fraud and compliance scenarios suggest a continuous effort to enhance its effectiveness and broaden its application within the financial industry.
Beyond the Headlines
The success of Jev highlights a broader shift in fraud detection strategies, moving towards specialized, high-speed AI models designed for specific tasks rather than relying solely on general-purpose LLMs or traditional rules-based systems. This specialization allows for greater accuracy and efficiency in environments where milliseconds and fractions of a cent matter. The emphasis on 'System One' thinking, characterized by fast and intuitive judgments, reflects a growing understanding of the need for AI to mimic human cognitive processes in real-time decision-making. This development could lead to a new generation of AI tools tailored for critical, high-volume applications, potentially setting a new standard for fraud prevention and compliance across various industries. The ability to automate complex risk checks with high accuracy also has implications for regulatory compliance, making it easier for financial institutions to meet stringent requirements.
