Underwriting Risk Analysis AI

FastAPI backend. Suggests APPROVE / REJECT / REFER_FOR_MANUAL_REVIEW on insurance proposals, with confidence % + template-based explanation (no LLM, deterministic).

No real data given -> synthetic dataset used

data/generate_synthetic_data.py builds 10k synthetic proposals using a hand-coded risk formula (age, smoker, BMI, credit score, claims history, occupation risk, etc)

• noise. Swap this for real historical proposal data when you get it -> same pipeline (train_model.py) retrains without other code changes, as long as column names match FEATURES list.

Pipeline

1. data/generate_synthetic_data.py -> underwriting_data.csv
2. ml/train_model.py -> trains RandomForestClassifier -> model.pkl + feature_meta.json (feature_meta.json stores feature_importances_ + risk thresholds, used by explanation engine)
3. app/ -> FastAPI serving layer
   • model_service.py: loads model, predicts, maps probability -> suggestion + confidence
   • explain.py: template engine, flags risky/positive features per applicant, ranks by model's feature importance, fills sentence templates
   • main.py: /api/v1/underwrite POST endpoint

Decision logic

• predict_proba(approve) > 0.58 -> APPROVE
• predict_proba(approve) < 0.42 -> REJECT
• in between -> REFER_FOR_MANUAL_REVIEW (model itself isn't confident -> don't let it decide)

confidence = how far probability sits from the decision boundary, scaled 0-100. risk_score = 100 - approve_probability*100 (0=safest, 100=riskiest).

Run

pip install -r requirements.txt
cd data && python generate_synthetic_data.py && cd ..
cd ml && python train_model.py && cd ..
uvicorn app.main:app --reload --port 8000

Docs: http://localhost:8000/docs Endpoint: POST http://localhost:8000/api/v1/underwrite

Sample request

{
  "age": 45, "annual_income": 800000, "sum_assured": 5000000,
  "bmi": 29.5, "smoker": 1, "alcohol_consumption": 1,
  "pre_existing_disease": 0, "family_medical_history": 1,
  "occupation_risk": 1, "credit_score": 610,
  "num_previous_claims": 1, "years_with_insurer": 0
}

When you get real data

1. Map your real columns to the FEATURES list in train_model.py (rename or extend).
2. Replace underwriting_data.csv with real historical approved/rejected proposals.
3. Rerun train_model.py -> new model.pkl auto picked up by API, zero other changes.
4. Update explain.py risk_thresholds to match real underwriting policy cutoffs (currently guessed: age 55, BMI 30, credit 650, claims >2).