Machine Learning · MLOps · 2025
Aurora: Real-Time Fraud & Anomaly Detection
A sub-15ms fraud detection pipeline processing 25,000 transactions/sec with a calibrated ensemble model, cutting chargebacks by 44%.
- Role
- Lead Data Science Engineer
- Timeline
- 12 weeks
- Year
- 2025
- Disciplines
- Machine Learning, MLOps
The Challenge
Modern payment systems face adversarial fraud rings that adapt their patterns within hours. Existing batch-scoring mechanisms had a 20-minute latency window, allowing fraudulent spikes to clear before risk flags were raised.
The engineering imperative was twofold: compute real-time velocity features across sliding 5-minute and 1-hour windows, and execute an ensemble inference model within a strict 15-millisecond SLA without blocking payment checkout.
Streaming Feature Architecture
We deployed a decoupled Kafka cluster ingesting raw transaction telemetry, feeding a low-latency Redis feature store with dual-write semantics. In-memory sliding window counters compute card velocity, IP entropy, and geographic displacement in under 3 milliseconds.
By formalizing schema contracts via Protocol Buffers and employing automated regression tests, we eliminated the classic divergence between offline training features and online inference features.
Model Architecture & Calibration
The scoring engine blends a LightGBM tree ensemble for tabular feature interactions with a PyTorch temporal attention network capturing transaction sequence dynamics. Model outputs are calibrated using isotonic regression to ensure predicted risk probabilities directly correspond to empirical default rates.
Inference was optimized using NVIDIA Triton Inference Server with dynamic batching and INT8 quantization, slashing GPU memory footprint by 65% and stabilizing p99 latency at 11.8ms.