12.4 Phase 4 — Advanced underwriting

Goal

Move from rule-based to model-based scoring while keeping rules as hard guardrails. Demonstrate ~10 – 20% improvement in approval-to-default trade-off.

Features

• Feature store (Feast or homegrown).
• Model registry (MLflow).
• Champion-challenger framework for production rollout.
• Gradient-boosting scoring model for approval-side.
• Decision explainability — SHAP per decision cached.
• Model monitoring — drift, calibration, fairness.
• Tally automation — vendor extractor live for top accounting software.
• E-invoice automation for invoice products.
• Anchor data automation for SCF programmes (one anchor live).
• Richer fraud signals — device fingerprint, consortium feed (if available).
• Periodic AA refresh for portfolio monitoring.

Integrations

• AA periodic re-pull workflow.
• Tally / Zoho parser maturity.
• Anchor ERP / accounting integrations.
• Device-fingerprint vendor.
• Consortium feed (if joined).

Team

• Add 2 – 3 data scientists.
• Add 1 ML engineer.
• Add 1 – 2 risk analyst (model monitoring).

Time

4 – 6 months.

Build vs buy

• Model itself: Build (XGBoost / LightGBM).
• Feature store: Feast (open source) or homegrown on warehouse.
• Model registry: MLflow.
• Monitoring: Build initially; evaluate Arize / Fiddler at scale.

Acceptance criteria

• Model in production with champion-challenger.
• Demonstrated improvement vs rule-based baseline.
• Explainability for every decision.
• Model monitoring dashboard live.
• Fairness checks documented.

Risks

• Model performance worse than rule-based (common in early ML rollouts).
• Drift detection lag — model decay unnoticed.
• Explainability gaps → FPC / compliance issues.

Go / no-go for Phase 5

• Model stable in production for ≥ 3 months.
• Risk team comfortable with model governance.