Tomás Ibáñez

Replaced a brittle static rules list with a real-time XGBoost and isolation-forest ensemble scoring roughly 4 million transactions daily. Engineered 220 velocity and graph features on a Kafka and Feast pipeline and served scores at a 38-millisecond p99 behind a FastAPI service. Cut net fraud losses by 43% while holding the false-positive rate under 1.8%.

Designed an unsupervised anomaly detection layer using autoencoders and isolation forests to catch account-takeover attempts that supervised rules missed. Combined device, behavioral, and session signals into a single risk score with human-in-the-loop review. Flagged 92% of confirmed takeover attempts at a 0.4% alert rate, protecting more than 300,000 active accounts.