Design Real-Time Fraud Detection — System Design Practice | AmanAI Lab

Sheet›Design Real-Time Fraud Detection

AmazonGoogleMSFTApple

45:00

Hard

The Problem

Problem

Design a real-time fraud detection system for a payment processor handling 10,000 transactions per second that must make an allow/deny decision within 100 milliseconds.

Requirements:

•Detect fraudulent credit card transactions in real time
•P99 decision latency ≤ 100ms
•False positive rate ≤ 0.1% (too many declines hurt customer experience)
•False negative rate ≤ 1% (missed fraud is costly)
•System must adapt to new fraud patterns within 24 hours

What you'll be assessed on

Real-time feature engineering, model architecture choices, latency budget, the FP/FN trade-off, and online learning for drift adaptation.

Scale & Constraints

▸10K transactions/second at peak
▸Decision latency ≤ 100ms P99
▸False positive rate ≤ 0.1%, false negative rate ≤ 1%
▸Fraud patterns change weekly — model must adapt
▸99.999% availability (payment infrastructure)
▸Regulatory requirement: store all decisions with explanations for 7 years

Must Cover0/10

Hints (if stuck)

💡 Velocity features (N transactions in last 1h/24h per card/merchant) are the most predictive.

💡 Rule engine first (fast, interpretable), then ML model for borderline cases — reduces latency.

💡 Latency budget: features from Redis ≤ 5ms, model inference ≤ 20ms, rules ≤ 5ms, response ≤ 70ms buffer.

💡 Keep a human review queue for high-confidence fraud (false negative cheaper to miss if caught later).

💡 SHAP values for explainability — required by financial regulators in many jurisdictions.

0 words · auto-saved

Back to sheet

Mock interviewTake quizFlashcards

Problem

Design a real-time fraud detection system for a payment processor handling 10,000 transactions per second that must make an allow/deny decision within 100 milliseconds.

Requirements:

•Detect fraudulent credit card transactions in real time

•P99 decision latency ≤ 100ms

•False positive rate ≤ 0.1% (too many declines hurt customer experience)

•False negative rate ≤ 1% (missed fraud is costly)

•System must adapt to new fraud patterns within 24 hours

What you'll be assessed on