AI Credit Underwriting for a Lending Startup
Real-time underwriting models that helped a Series A lending startup approve more good borrowers without raising loss rates.
Key Details
| Challenge | Manual credit reviews capped growth; rules-only scoring rejected thin-file applicants the startup wanted to serve. |
|---|---|
| Solution | A hybrid ML underwriting stack with explainable scores, policy overrides and an analyst review queue. |
| Technologies | Python, XGBoost, FastAPI, PostgreSQL, AWS |
Technologies used
Python
AWS
PostgreSQL
FastAPI
Docker
Redis
Client background
A Series A consumer-lending startup was growing application volume faster than its underwriting team. Spreadsheet rules and bureau cutoffs either approved too aggressively or blocked thin-file customers that were core to the product thesis.
Key challenges
- Analysts could not keep up with peak application spikes without adding headcount.
- Thin-file and new-to-credit applicants were over-rejected by rigid bureau rules.
- Founders needed explainability for investors, partners and regulators.
- Shadow models in notebooks never made it into the live decision path.
What we built
- Feature store for bureau, bank-transaction and application signals with freshness checks.
- Gradient-boosted underwriting model with SHAP-style driver explanations for analysts.
- FastAPI decision service with policy overlays and human review for edge cases.
- Monitoring for approval rate, loss early-warning and score drift.
Project team: 5 engineers across AI/ML, backend and product engineering — delivery over 12 weeks for a startup team that needed to ship, not slide decks.
How we delivered
Scope
Aligned risk appetite, excluded populations and what “good approval” meant for the round.
Data
Joined bureau, KYC and transaction feeds into a decision-ready feature layer.
Model
Trained and calibrated scores with holdout and time-based validation.
Ship
Cut over behind feature flags with analyst override and weekly risk reviews.