
GitHub
ML engineer building fraud detection systems that block millions in losses daily. Power of 2.

Cut fraud false positives from 2.3% to 0.8%, recovering $31M/month
Stripe Radar's fraud model had a 2.3% false positive rate, blocking $47M/month in legitimate transactions. The model retrained weekly, meaning new fraud patterns took 7-14 days to catch. Merchants were churning because legitimate customers were getting blocked.
Built an adaptive model architecture that combines a base XGBoost model with a real-time neural network that learns from merchant-specific feedback signals. Created a streaming feature store using Kafka + Redis that computes 200+ features in under 10ms. Implemented merchant-specific threshold tuning using Bayesian optimization.
False positive rate dropped from 2.3% to 0.8%, recovering $31M/month in previously blocked legitimate transactions. New fraud pattern detection time went from 7 days to 4 hours. Model serves decisions at 50K transactions/second with P99 latency under 15ms.

PayPal's fraud ops team manually reviewed 50K flagged transactions daily, but only 8% were actual fraud. Analysts spent 80% of their time on false alarms.
Built an anomaly clustering system that groups related suspicious transactions and prioritizes review by estimated loss exposure. Used Claude to generate human-readable alert summaries.
Analyst efficiency improved 4x. True positive rate in reviewed alerts went from 8% to 34%. Average fraud case resolution time dropped from 72 hours to 6 hours.