Bloomberg ML System Design: Real-time Fraud Detection

You are asked to design a low-latency, production-ready machine learning system that detects fraudulent e-commerce transactions in real time.

The core task is to ingest streams of transaction events (user profile, payment method, device, location, purchase history), compute real-time features, run an online inference model, and output a fraud probability and binary decision within strict latency bounds (<=100 ms). You must also provide data storage for auditing and retraining, a safe deployment strategy for model updates, and monitoring for model quality and data drift.

High-level flow you should discuss:

• Data ingestion and validation (streaming platforms, gateways)
• Real-time feature computation (feature store, windowing, sessionization)
• Low-latency model serving (lightweight ensemble or neural network optimized for inference)
• Decisioning and alerting (rules + score thresholds, manual review queue)
• Storage and offline pipelines (cold store for full transactions, warm store for recent context)

Skill signals interviewers expect:

• Knowledge of online inference and feature stores (latency, consistency, cold starts)
• Trade-offs between precision/recall and user experience (thresholding, cost of false positives)
• Scalability and reliability design (sharding, autoscaling, failover)
• Monitoring and observability (latency SLOs, data/model drift detection, alerting)

Be ready to justify design trade-offs, propose A/B or shadow testing strategies, and describe responses to evolving fraud patterns.