Fighting Fraud at Scale: Insights from building a real-time rules engine

DoorDash built a real-time rules engine to decouple fraud decision logic from application deployments. The system uses checkpoints as entry points, models context as a DAG of reusable facts, and supports shadow evaluation, backtesting, and controlled experimentation before rules go live. Running at ~10,000 RPS with ~15,000 facts, it reduced incident mitigation time by over 80% and cut rule deployment turnaround from days to minutes. The platform is now expanding beyond fraud into Trust & Safety, Logistics, and other domains as a general-purpose decisioning engine.