AI-Based Fraud & Anomaly Detection Model for an Online Payments Company

Overview

Built a machine learning model for an online payments company that flags transactions showing patterns associated with fraud, based on velocity, device fingerprinting signals, and deviation from a customer's typical behavior. Flagged transactions are routed to a manual review queue rather than being automatically declined, keeping a human decision-maker in the loop for anything the model isn't highly confident about. This gave the company's small trust-and-safety team a way to focus manual review time on the transactions most likely to be fraudulent, instead of either reviewing everything or relying on static rule-based thresholds that were increasingly easy to work around. The model was trained on the company's own historical transaction and confirmed-fraud data and is retrained on a regular cadence as fraud patterns evolve. An online payments company's small trust-and-safety team was either reviewing every transaction manually or relying on static rule-based thresholds that were increasingly easy for fraudsters to work around. We built a machine learning model flagging transactions showing fraud-associated patterns — velocity, device fingerprinting, behavioral deviation — routing flagged transactions to manual review rather than automatically declining them, keeping a human decision-maker in the loop. We trained the model on the company's own historical transaction and confirmed-fraud data, validating flagged results against known past fraud cases before it touched live transactions. It ran in parallel with the existing rule-based system for several weeks, comparing outcomes, before becoming the primary detection method. Confirmed fraud losses dropped by an estimated 25%, while the volume of legitimate transactions incorrectly flagged for manual review also fell — improving accuracy in both directions rather than trading one for the other.