Compliance & fairness

Every decision decomposes into ranked reason codes. A rejection is issued with the exact, plain-language factors that drove it — and what would improve them — satisfying the requirement to give borrowers a clear reason and a path, not a black-box “no”.

The model uses only these business-behaviour signals. None is gender, caste, religion, age, or location:

• Healthy inflow-to-outflow ratio· cashflow
• Stable monthly cash flows· cashflow
• Strong post-obligation surplus· repayment
• Comfortable balance vs. EMI· repayment
• Established operating vintage· stability
• Regular EPFO contributions· stability
• Timely GST filings· compliance
• Consistent GST filing· compliance
• Low EMI burden vs. inflows· leverage
• Few payment bounces· leverage
• Growing turnover· growth
• Growing UPI volume· growth
• Rising power consumption (activity)· utility
• Utility bills paid on time· utility

The production model is a transparent WOE/logistic scorecard; every bin, weight and point is versioned in git and exportable (Model & Weights page). Any historical decision can be reconstructed exactly.

Data is pulled only with explicit, purpose-bound consent via the Account Aggregator (see AA Consent), limited to the fields the seven dimensions need, for a single underwriting purpose and a bounded time window.

Expected loss is computed as PD × LGD × exposure, with LGD held at 40% (a conservative unsecured-MSME assumption). Rates scale with PD so expanded approvals stay priced for their risk.