ML Engineer — Fraud Detection Model

500k transactions/day. 2.3% false positive rate on fraud blocks. Every false positive is a declined legitimate transaction — real revenue, real customer friction. Our current system is a rules engine built in 2018. It's not learning, it can't adapt, and adding a new rule requires a 3-week change management process. We need a proper ML model. Constraints you need to know upfront: inference must run in under 30ms (we're latency-sensitive), the model needs to be interpretable enough that our compliance team can explain individual decisions, and class imbalance is severe (~0.1% fraud rate). We have 18 months of labelled transaction data, ~150M rows. Our internal team handles deployment — we need the trained model and clear documentation.