Building a Real-Time Fraud Engine for Zelle Payments

In financial services, the ability to spot fraud faster can be a key differentiator.

At Data Summit 2018, Karun Komirishetty, senior manager, Software Engineering, Capital One, presented a session on how Capital One has created a fraud engine for Zelle payments, which are replacing traditional forms of payments such as checks, cash, and credit cards.

Zelle is a U.S.-based digital payment network owned by Early Warning System (EWS) that provides a convenient way to exchange money with users who are also part of the Zelle network. Users send money using a recipient’s email address or phone number.

The talk covered Capital One’s move from a micro-services-based fraud detection system to a new system that relies on stream processing (Apache Flink) and machine-learning to detect fraud.

According to Komirishetty, the downsides of the services-based architecture led the company to look into an alternative. Downsides included the fraud engine having to make multiple backend calls; the fact that resource utilization was high on systems of record; and new data sources increasing latency. In addition, it was not really high performance and not learning from historical fraud scoring requests.

The benefits of the new system are in performance, with real-time fraud scoring taking less than 10ms; reliability, with a reduction of the load on system of record, and external system issues isolated from the customer experience; feedback and learning, with scoring response feedback able to be analyzed by DAs, and the ability to improve the model based on historical data; and easy monitoring and trouble-shooting of issues.

Many presentations from Data Summit 2018 have been made available for review at