Sentinel

01The problem

Fraud detection has to score events in real time and explain itself. A black-box score is not enough when a declined transaction gets reviewed.

02Approach

• 01Built an inference workflow covering feature preparation, model serving, explainability, and monitoring.
• 02Served predictions through FastAPI with scikit-learn, XGBoost, and LightGBM models.
• 03Used Redis and PostgreSQL for event handling, caching, and persistence.
• 04Added explainability so each decision can be traced and audited.

03Outcomes

• Runs inference in real time with production-oriented ML operations.
• Every score ships with an explanation a reviewer can inspect.
• Feature, model, and monitoring concerns stay separated in the service.