Design real-time fraud detection under 50ms
Company: NVIDIA
Role: Software Engineer
Category: ML System Design
Difficulty: easy
Interview Round: Technical Screen
Design a real-time fraud detection system for a payments company that processes millions of transactions per day.
Requirements:
- For each incoming transaction, the system must decide **Approve / Flag / Block**.
- End-to-end decision latency must be **≤ 50 ms** per transaction.
- Sustain **10,000+ requests/second** (RPS) and tolerate promotional spikes (e.g., Black Friday) with high transaction success rate.
- The ML model(s) must be updatable **without downtime** (no service interruption during model rollout).
Describe the architecture, data/feature flow, model serving strategy, scaling and reliability approach, and how you would operate/monitor the system in production.
Overview: This question evaluates expertise in designing low-latency, high-throughput ML-powered systems for online decisioning, covering competencies in real-time inference, feature pipelines, model serving, scaling, and operational reliability within the ML System Design domain.