Design a Fraud Detection System

Intermediate 60 minPhonePeRazorpayPaytmStripeAmazonAmerican Express

Design a real-time fraud detection system for a payments app that decides whether to approve, challenge (OTP) or block each transaction.

Fraud is rare, labels arrive weeks later (chargebacks), and fraudsters change tactics quickly.

Functional requirements

Hint 1Sign in to use hints
Labels show up 30–90 days later. How does that affect training and evaluation?
Hint 2Sign in to use hints
The most useful features are aggregates: transactions in the last 10 minutes, new device, distance from usual location. Where are they computed?
Hint 3Sign in to use hints
Streaming feature pipeline (Flink) into an online feature store (velocity counts, device/IP reputation, graph features like shared devices) → gradient-boosted model + rules → decision service with thresholds per risk band; delayed labels from chargebacks plus analyst reviews; time-based train/test splits; monitor fraud rate, approval rate and drift; shadow mode for new models.

What the interviewer expects in each area unlocks after your first submission.

Write requirements and estimates, draw the architecture, and get your design scored. Your drafts save as you type and follow you across devices.