Inspiration

About 190 million Indian adults have no usable credit score. Lenders decide from bureau history, so gig workers, shopkeepers, farmers, students and homemakers are refused or pushed to instant-loan apps, even when their bank account shows steady income and on-time bills. The evidence is already in their bank statements, but lenders can't use it: statements are thousands of cryptic UPI narrations, they are easy to forge, and there is no explainable way to turn cash flow into a decision.

What it does

CreditBridge turns 6-12 months of bank data into an explainable credit decision in seconds.

  • Connect data: upload a bank statement (PDF/CSV), connect through an RBI Account Aggregator consent screen, or try a sample profile.
  • Authenticity checks: running-balance continuity (catches edited amounts), PDF metadata edits, duplicate transaction IDs, date gaps, salary consistency, and round-tripping (money sent back to the same person within 72 hours is fake income and is excluded). A failed check blocks auto-approval and shows the evidence rows.
  • Transaction understanding: an ML categorizer reads every narration into 27 categories.
  • Decision: a 300-900 score, A-E band, approve / review / decline, reason codes, FOIR-based affordable EMI vs the requested EMI, and a recommended loan amount.
  • How to improve: concrete actions with the projected score after each one.
  • For borrowers: summaries in English, Hindi and Telugu, plus a printable decision letter.
  • For lenders: a portfolio console with flagged statements, and a model transparency page for risk and compliance teams.

How we built it

  • Simulator (Python): no public dataset of Indian bank statements with repayment outcomes exists, so we built a persona-based simulator: salaried, gig workers, shopkeepers, farmers, students and homemakers, with narration formats from 8 banks, truncation and typos. Hidden traits (money discipline, income-shock risk, liquidity stress, loan-app and betting propensity) drive behaviour. Default is simulated over a separate future window with fresh shocks and noise, so it can't be read straight off the statement. 6,000 applicants, 2.85M transactions.
  • Model 1, transaction categorizer: char 3-5-gram + word TF-IDF → multinomial logistic regression (scikit-learn), evaluated on merchants never seen in training.
  • Model 2, risk scorecard: 17 cash-flow features (bill regularity, savings rate, FOIR, loan-app use, bounces, balance before payday, betting, income variability...) → a sign-constrained L2 logistic regression fit with SciPy L-BFGS-B: every weight is forced in the direction lending logic expects or to zero, so a reason code can never be reversed by collinearity. Selection and scaling are refit inside each of 5 CV folds, and it is benchmarked against a WoE scorecard and gradient boosting on the same features and folds. Every decision splits exactly into per-feature contributions, which become the reason codes and drive the counterfactual search.
  • Policy guardrails: statements under 90 days or 30 transactions are never auto-approved; no verifiable income is declined; failed authenticity checks block approval. Approve/review PD cut-offs are fixed policy, not tuned on outcomes.
  • No LLM or generative AI anywhere in the product: the multilingual summaries are templates filled with the applicant's numbers.
  • Runtime: models exported to JSON and run with numpy only (parity with scikit-learn asserted at training time), served by FastAPI as a Vercel Python function. pdfplumber parses statements. Nothing is stored.
  • Frontend: Next.js 16, React 19, TypeScript, Tailwind 4, Recharts.
  • Quality: 107 pytest tests (parsers, forgery checks, scoring invariants, API) and GitHub Actions CI on every push.

Results

Component Metric
Categorizer, unseen merchants accuracy 0.860, macro-F1 0.797
Risk scorecard (deployed, 5-fold CV) AUC 0.718, KS 0.334, Gini 0.436, Brier 0.146
Benchmarks (same features and folds) WoE scorecard 0.726, gradient boosting 0.723
Bureau-like baseline (EMI + any bounce) AUC 0.609
Default rate by band A → E 3.7% → 7.1% → 13.1% → 25.5% → 44.1%
Auto-approved / declined default rate 6.6% / 35.4%

Real-data validation (Statlog German Credit, 1,000 real applicants): alternative features alone reach AUC 0.751 vs 0.625 for bureau features. On the 89 applicants with no or fully-paid prior credit, alternative data reaches 0.682 while bureau features carry no usable signal (0.338); that subset is small and defined by a bureau field, so we treat it as directional.

Fairness: gender is never a model input. Approval rate is 22.8% for women vs 21.0% for men; among people who repaid, 25.7% vs 25.0%.

Challenges we ran into

  • No real labelled data: we built a behavioural simulator where default depends on hidden traits, not on observed columns.
  • Explainability vs accuracy: the black-box benchmarks were less than 0.01 AUC better, so we shipped the transparent scorecard. A plain logistic model flipped signs on correlated features, so we constrained every coefficient's direction instead of dropping features by hand.
  • Fraud: without authenticity checks a cash-flow model is easy to game, so integrity runs before scoring and can block approval, while tolerating genuine quirks (missing balances, PDFs re-saved to remove passwords, small transfers between friends).
  • Serverless limits: the runtime uses only numpy, with scikit-learn parity asserted at training time.

Accomplishments that we're proud of

An end-to-end, deployed underwriting pipeline that reads raw statements in three formats, catches forged PDFs and round-tripped income, and explains every decision in three Indian languages.

What we learned

For thin-file borrowers, bureau data carries little signal. Cash-flow behaviour (bill regularity, savings, FOIR, buffers) carries most of it, and authenticity checks are what make cash-flow lending safe.

What's next for CreditBridge

  • Re-fit on real repayment outcomes with an NBFC partner, using live Account Aggregator data
  • OCR for scanned statements
  • Voice explanations for low-literacy users
  • Monitoring for drift and fairness on real outcomes

Prototype for research only, not a lending decision. The risk model is trained on simulated statements.

Built With

Share this project:

Updates

Submission history