Confidential Client - Fraud Risk Scoring (KSA).

The platform needed instant onboarding decisions while also maintaining a separate continuous-monitoring view of account behavior over time.

The behavioral model had to leverage both unlabeled anomaly signals and historical confirmed fraud without exploding false positives on sparse or shifting populations.

I built a real-time onboarding scorer using customer profile, document, device, network, location, channel, and early behavioral signals to return a risk score and three-way decision into the middleware layer.

For ongoing risk monitoring, I designed a semi-supervised architecture where an unsupervised ensemble first measured account abnormality, then fed that continuous anomaly score into an advanced gradient-boosting supervised classifier alongside 50+ engineered features derived across micro, short, medium, and baseline windows.

To keep the system reliable in production, I added insufficient-data routing, cohort-aware standardization, distribution-based drift checks, post-training probability calibration, and feature-contribution explanations for elevated-risk accounts before both models were exposed to the fraud-management platform through service middleware.