Inspiration
India's online sellers and freelancers can have real payment histories but still struggle to explain their cash flow or evaluate borrowing offers. We built Paypilot AI to turn a seller's PayPal transaction history into understandable financial signals, while keeping the borrower in control of every money-related action.
What it does
Paypilot AI is a cash-flow credit copilot for India's online sellers and freelancers. It reads 12 months of PayPal income, produces an explainable 300-850 PayScore, and shows the drivers behind it with TreeSHAP.
The app also includes:
- Offer Guard to compare an advertised loan rate with the true annualized cost using the cash-flow IRR, including processing fees.
- Affordability checks and stress scenarios for proposed borrowing.
- An AI copilot that answers questions such as "Can I afford ₹50k?" and presents a separate approval step before a sandbox payout can be requested.
- A revocable Income Passport with a public verification page, plus tools for negotiation, repayment, revenue volatility, and income-by-client views.
For a safe judge experience, the app provides synthetic demo personas and PayPal sandbox support. It is not a real lending product or a real credit decision system.
How we built it
The frontend is a React, TypeScript, Vite application styled with Tailwind CSS, with charts and data tables for the financial workspace. The backend is a FastAPI/Python service.
For scoring, we use a LightGBM model trained on synthetic development data. A shared feature pipeline converts income and profile inputs into a 300-850 score, and SHAP explains the model's contribution drivers. Counterfactual suggestions are rescored rather than guessed. Offer Guard derives borrower cash flows from principal, fees, and instalments, then solves the monthly IRR to reveal effective annual cost.
PayPal sandbox sign-in uses OAuth with PKCE. The linked account token lives only in an encrypted, HttpOnly cookie in that browser. The app reads only the linked account's permitted sandbox history, and switching back to a demo persona unlinks it.
Challenges we ran into
The core challenge was making finance features useful without overstating what they mean. We made the model's synthetic-data limitation clear, labeled score explanations as model attributions rather than causal claims, and made affordability outputs estimates rather than regulated advice. We also designed the demo so every payout path requires explicit user approval and stays in sandbox or simulated mode.
We also had to make offer analysis robust: uploaded PDFs are size- and page-limited, parsed locally, and ambiguous values are not treated as verified. For privacy, a linked PayPal account is isolated to the browser session so data from one visitor cannot appear for another.
What we learned
We learned that trustworthy financial AI needs more than a prediction. Users need clear explanations, honest limitations, meaningful comparisons, privacy boundaries, and control over actions. Paypilot AI combines those ideas in one workflow: understand income, inspect a score, question a loan offer, test affordability, and approve any sandbox payment only when the user is ready.
What's next
We would validate the system on appropriately governed real-world data, conduct fairness and calibration work, strengthen lender-directory coverage, and continue expanding safe, permissioned PayPal integrations.
Built With
- fastapi
- lightgbm
- oauth
- paypal
- python
- react
- render
- scikit-learn
- shap
- tailwind
- typescript
- vercel
- vite
