Design E-Commerce Fraud Detection

Problem Statement: E-Commerce Fraud Detection Platform

E-commerce fraud detection is a real-time risk-scoring system that sits in the checkout path and decides, in under 80 milliseconds, whether to approve, challenge, or block a transaction — balancing two opposing costs: the chargeback loss from approving fraud, and the lost revenue from declining a good customer. Stripe Radar, PayPal, and Sift all live on this knife-edge, and the interviewer is listening for whether you frame it economically (basis points of fraud versus false-decline revenue loss) rather than hand-waving 'use AI'.

The threats are specific and each shapes the design. Card testing: bots burn thousands of micro-authorizations per hour to validate stolen card numbers, which means velocity counters must be both merchant-scoped and instrument-scoped — a global counter would either miss the attack or flag a legitimate flash-sale power user as collateral damage. Account takeover presents as a familiar device with a new shipping address buying high-value digital goods. Promo and refund abuse is e-commerce-specific. And chargeback losses are the bottom line the whole system optimizes against.

The architectural stance to state early is the consistency split. Velocity counters and device-reputation caches are AP — a read that is a few milliseconds stale is fine, and we never want the fraud check to be the thing that takes down checkout. Decision records and case state are CP — auditors and disputes require linearizable, immutable records of what we decided and why. Tie every box you draw to a measurable outcome: authorization fraud rate, review-queue depth, analyst handle time, and model precision at a fixed recall.

The failure that defines the design: the ML model serving times out during a Black Friday peak. A naive system blocks (killing revenue) or silently approves (inviting fraud); the correct system degrades to rules-only with an elevated challenge rate, because a fraud platform must always return a defensible decision, never an error, on the money path.