Real-time fraud detection

Design a real-time fraud-detection system for a payments platform processing 5k transactions/sec, where each transaction must get a risk score within a 100ms p99 budget to allow/deny/challenge. Fraud is rare (<0.5%), labels arrive days later via chargebacks, and adversaries adapt continuously. Cover the feature pipeline (including velocity/aggregate features), the model and decision layer, and how you keep up with evolving fraud patterns.