Leakage-safe temporal evaluation and point-in-time feature design.
Real-Time Fraud Detection Platform
A production-style fraud platform that turns payment events into calibrated risk scores, policy decisions, alerts, drift reports, and analyst-facing operational views.
Built an end-to-end system that separates model risk from business policy while connecting event streaming, point-in-time features, explainability, monitoring, and review workflows.
01 / PROBLEM
What the system is trying to solve.
A useful fraud system needs more than a classifier. It has to score events reliably, preserve event-time correctness, expose why a decision happened, and continue behaving predictably as data changes.
02 / SYSTEM
How it works.
- 01
Ingest payment events through a Kafka-compatible Redpanda stream.
- 02
Build leakage-safe chronological features and score events through a FastAPI inference service.
- 03
Apply APPROVE, REVIEW, and BLOCK policy thresholds separately from model probabilities.
- 04
Persist decisions through idempotent consumers and a transactional outbox pattern.
- 05
Track delayed-label performance, SHAP explanations, drift statistics, and champion/challenger retraining signals.
03 / ENGINEERING
Decisions that shaped the project.
XGBoost model experimentation with MLflow tracking and Optuna tuning.
SHAP-backed explanation and alerting workflows.
PSI, KS, and JS drift monitoring plus delayed-label metrics.
Next.js command center with operational views and Playwright end-to-end coverage.
04 / EVIDENCE
A few concrete signals.
0.9266
Precision0.3108
Recall0.4654
F10.3363
PR-AUC05 / LINKS