Captured from the live deployment, not a mockup
01 · The problem
Built for the lenders that actually underwrite thin-file borrowers, NBFCs, microfinance institutions and BNPL platforms, and shipped it as a pip-installable library they embed rather than another dashboard to stand up.
02 · How it works
- 01
Made every decision explainable by construction, with SHAP attributions doubling as regulator-style adverse-action reasons, and currency and policy configurable per market.
- 02
Validated against 150,000 real borrowers with real default outcomes rather than the synthetic data it was developed on, reaching 0.82 AUC on half the feature set.
- 03
Thin-file borrowers have no bureau history, so scored them from behavioural signals instead: income entropy, rent consistency, bounce history and EMI burden, with LightGBM handling the sparsity natively.
- 04
Detected life events straight from cash flow with ruptures PELT change-point detection, surfacing EMI closures and income step-ups a static snapshot would miss.
03 · What it cost, and what it returned
Validated a credit model against 150,000 real borrowers with real default outcomes, deliberately closing the circularity gap of evaluating on the synthetic data it was built with.