Design real-time fraud scoring

Payments platform. 4,000 transactions per second at peak, 800 average. Every transaction must receive a fraud score before authorisation, and the end-to-end budget for scoring is 50ms at p99.

Fraud labels arrive from chargebacks, 30 to 90 days after the transaction. Roughly 0.15% of transactions are fraudulent. A false negative costs the average transaction value, about £70. A false positive blocks a legitimate customer, and the business estimates the lifetime cost of that at roughly £15 — but it is politically far more visible.

Regulatory requirement: any declined transaction must be explainable to the customer on request, and the decision must be reproducible for two years.

Design this. I want the serving architecture, the feature strategy, the model choice, the threshold decision, and the monitoring — and for each, what you are optimising for and what you are giving up.