THE PIPELINE
From raw data to credit decision in milliseconds.
A walkthrough of the TrueScore scoring architecture across data ingestion, feature engineering, model scoring, explainability, and delivery.
Step 1 - Data Ingestion
Scaffold coverage includes bureau pulls, AA consent and fetch, alternative data via Perfios / Finbox, DigiLocker verification, GSTN fetches, and normalisation before scoring.
Step 2 - Feature Engineering
Individual pipeline
| Feature | Source | Type |
|---|---|---|
| FOIR | Bank statement / AA | Ratio |
| DTI | Bureau + income docs | Ratio |
| Repayment consistency score | Bureau history | Derived |
| Cash flow regularity | AA / bank stmt | Derived |
| Bureau vintage | CIBIL / Experian | Raw |
| Enquiry intensity (90d) | Bureau | Count |
Business pipeline
| Feature | Source | Type |
|---|---|---|
| DSCR | Bank stmt / Perfios | Ratio |
| GST turnover trend (12m) | GSTN | Trend |
| Filing compliance rate | GSTN | Rate |
| Promoter credit score | Bureau | Raw |
| Charge or lien on assets | MCA21 | Flag |
| Sector risk overlay | RBI sectoral data | Score |
Step 3 - LightGBM Scoring Engine
Prompt 02 frames LightGBM as the preferred model family for speed, categorical feature handling, and better explainability tradeoffs than deeper black-box alternatives.
| Model | AUC-ROC | KS Stat | Gini |
|---|---|---|---|
| Logistic Regression (baseline) | 0.74 | 0.38 | 0.48 |
| Random Forest | 0.79 | 0.44 | 0.58 |
| XGBoost | 0.82 | 0.48 | 0.64 |
| TrueScore (LightGBM) | 0.91 | 0.62 | 0.82 |
Step 4 - SHAP Explainability
Every score should map back to human-readable drivers so a lender can explain adverse action without treating interpretability as an afterthought.
Step 5 - Delivery
REST API, webhook delivery, batch scoring, and dashboard consumption should all resolve from the same scoring contract and audit trail.