kaustubspatil/Jarvis-underwriting-desk

JARVIS Underwriting Desk · Stark Financial

A live credit decisioning demo. Enter an applicant, and the app scores them with the real trained pipeline, prices the decision in money, explains it with SHAP, and writes it up in plain English.

The governance architecture

the MODEL decides real trained pipeline, deterministic SHAP explains exact, additive, auditable the LLM only narrates never changes a decision, and is optional

Configuration

Set OPENROUTER_API_KEY as a Secret (Settings → Variables and secrets) to enable AI narration. Optionally set OPENROUTER_MODEL.

Without a key the app falls back to deterministic wording. The decision, the money, the reason codes and the audit record are byte-identical either way, so the demo cannot break in front of an audience.

Model

credit_scoring_pipeline_nogender.pkl — a scikit-learn HistGradientBoostingClassifier trained on the Home Credit Default Risk dataset, with gender excluded from the feature set. Scored against a sealed 15% test split.