Build features visually, train hybrid scorecard + ML models, and validate them with the instruments credit risk actually uses — nothing reaches production without its evidence attached.
In commercial engagement with South Indian Bank and Jana Small Finance Bank.
Illustrative canvas · risk teams assemble this visually on the canvas.
The Ind AS 109 ECL transition is live — concluding FY 2026–27. Most banks are still reconciling origination and provisioning data in Excel. PramanaRiskAI eliminates that gap: the ECL model's population is identical to the origination population, because both draw from the same entity definitions — no feature drift between training and production scoring. Built for RBI model risk governance and DPDP data residency →
Generic ML platforms make your analysts re-teach credit risk to every tool. Pramana ships with the instruments a scorecard team actually reaches for — every instrument, built in.
Cohort delinquency curves by booking month — spot a bad vintage before it seasons.
Who is actually applying vs. who you approve — see population shifts before models drift.
Not mean-imputation everywhere — strategies that respect what each field means in credit.
Supervised binning with monotonicity checks — interpretable scorecards regulators can read.
Gini, KS, PSI, calibration, ROC — the full panel, computed on every train and every refresh.
Population and score drift tracked continuously — alerts before performance decays, not after.
Visual pipelines feed hybrid WoE + ML training — segments, samples, and target definitions versioned together.
Discrimination, calibration, stability, and TTD checks — the validation pack writes itself as you work.
Push to NirnayaDecisionAI for real-time verdicts or Pragna scoring jobs for batch — same lineage either way.
Drift, PSI, and performance tracked in production; champion-challenger refreshes close the loop.
Every model carries its evidence: input schema, WoE bins, score cut-offs, diagnostics, documentation, and governance state — the MRM pack your validators and RBI examiners actually want.
Watch a feature mart, a hybrid model, and its validation pack come together on your data — in weeks, not quarters.
