My role
Data scientist — anomaly detection and streaming inference.
Overview
Built a real-time scoring service on a streaming feature store. Combined a supervised gradient-boosted model with an unsupervised isolation-forest signal to catch both known and novel fraud patterns, and tuned the decision threshold against the cost of false declines. Deployed on a Kafka-fed pipeline with online features computed in Flink, and set up a champion/challenger harness so new models are proven on live traffic before promotion.
Role & impact
Impact
Increased fraud recall by 34% while holding the false-positive rate flat, saving ~$2.6M in annual losses.