02Credit Risk · ML · Decision Intelligence
Sentinel Credit.
An end-to-end BNPL risk intelligence platform for thin-file customers, combining modelling, monitoring, explainability and governance.
01 · Problem
Thin-file lending creates a decision problem, not just a modelling problem.
Customers with limited bureau history can be difficult to assess using conventional signals. The project explores how behavioural, affordability and repayment evidence can be combined without hiding uncertainty.
02 · Approach
Model the customer journey, then engineer the decision.
A synthetic BNPL environment was designed around customer, application and repayment grains, followed by feature engineering, challenger models, affordability logic, explainability, monitoring and governance controls.
03 · Decisions
Designed for trade-offs, not demos.
Natural keys and shared SQL features were chosen to preserve consistency across training and inference. Thresholds are treated as business decisions rather than defaulting to 0.5, and model outputs are separated from policy rules.
04 · Evidence
Performance is only useful when it survives deployment.
The evaluation plan includes discrimination, calibration, temporal validation, drift monitoring, fairness checks and challenger comparison. Thresholds are linked to affordability and risk appetite rather than selected in isolation.
05 · Reliability
Trust is part of the architecture.
The design includes human approval points, SHAP-based explanations, drift triggers, lineage and monitoring. Synthetic data is clearly separated from real customer evidence to avoid false claims of production performance.
06 · Reflection
Better simulation changed the whole project.
The data generator evolved into a hazard-based state machine because simplistic synthetic defaults produced unrealistic patterns. That redesign improved the credibility of both model development and monitoring.
System architecture