Creating a Real-Time Fraud Signals Platform
Specialized Engineering Scope & Mission Challenge
An international digital retail bank experienced escalating losses from account takeover (ATO) and authorized push payment (APP) fraud. Existing fraud scoring relied on legacy batch rules that ran minutes or hours after transaction completion. To stop fraudulent transfers before authorization, the bank needed an ultra-low-latency feature engineering and scoring pipeline evaluating device telemetry, behavioral biometrics, recipient velocity, and historical risk signals in under 150 milliseconds.
Granular Competency & Technical Requirement Matrix
Critical Engineering & Integration Complexities
- Strict 150ms total SLA budget including network round-trip, feature aggregation, model inference, and rule arbitration.
- Severe class imbalance where fraudulent events represent less than 0.01% of high-volume transaction traffic.
- Dynamic adversarial behavior requiring weekly model retraining and fast deployment without service interruption.
- Consumer protection regulations requiring clear explainability for any blocked transaction to prevent unjust financial exclusion.
Engineered Delivery & Execution Workflow
Feature Store Architecture
Designed an in-memory streaming feature store aggregating sliding-window transaction velocities in real time.
Real-Time Ingestion Pipeline
Connected mobile app signals, IP geolocation, and transaction metadata via Kafka event topics.
Low-Latency Inference Runtime
Exported gradient-boosted tree and neural network models into C++ ONNX runtimes for sub-10ms evaluation.
Multi-Tier Arbitration Logic
Implemented fallback decision ladders blending hard compliance filters with dynamic probability thresholds.
Dark Pipeline Replay
Backtested new model candidates against historical transaction streams to measure precision-recall without customer impact.
Analyst Case Console
Built an intuitive risk operations console displaying feature contribution waterfalls and recommended mitigation actions.
Canary Model Deployment
Rolled out updated scoring models via shadow canary routing, monitoring classification drift and false-positive spikes.
Production Toolchains, Protocols & Frameworks
06 | PROJECT TARGETS
Project Targets and Business Impact
Customer Confidentiality Note: Customer identity, proprietary codebases, and commercial program specifics have been anonymized in accordance with non-disclosure agreements.