How do you implement AI-powered fraud detection systems?

Fraud detection uses ML to identify suspicious transactions and activities in real-time. Architecture: data ingestion (real-time transaction stream from payment systems), feature computation (real-time feature engineering from transaction and historical data), model inference (score each transaction for fraud probability), decision engine (approve, decline, or flag for review based on score and rules), and case management (human investigators review flagged transactions). Feature engineering: transaction features (amount, merchant category, transaction type, time of day), velocity features (count and sum of transactions in last 1h/24h/7d), behavioral features (deviation from customer's normal patterns: unusual merchant, location, amount), device features (IP geolocation, device fingerprint, browser characteristics), and network features (connection to known fraud rings via shared attributes). Model approach: (1) Supervised: train on labeled fraud/not-fraud data using gradient boosting (XGBoost) or neural networks. Challenge: extreme class imbalance (fraud is typically 0.1-1% of transactions). Handle with SMOTE, class weights, or anomaly detection. (2) Unsupervised: isolation forest or autoencoders detect anomalous transactions without labeled data, useful for detecting novel fraud patterns. (3) Graph-based: build transaction networks and detect fraudulent communities using graph neural networks or community detection. Real-time requirements: score within 50-100ms for payment authorization, implement feature store for pre-computed features (Feast, Tecton), and use model serving infrastructure (Seldon, KServe) optimized for low-latency inference. Rule engine: supplement ML with rules for known fraud patterns (list of compromised cards, blocked countries, velocity limits), implement rule versioning and A/B testing, and allow fraud analysts to create and modify rules without developer involvement. Metrics: fraud detection rate (percentage of fraud caught), false positive rate (legitimate transactions declined), precision at top-k (accuracy of highest-scored transactions), and financial loss prevented.