Inspiration
I moved from Kigali to Miami in 2025. Back home I had a bank account, mobile money and years of rent paid on time. In the US, none of it existed: no credit score, so no apartment, no card, no loan. Millions of immigrants and international students arrive the same way, financially invisible despite a real track record.
What it does
Credit Virgin turns the paper trail newcomers already have into a score a US landlord can trust.
- Upload anything, in any language. Mobile money statements, bank statements, rent receipts, as PDFs or photos. Gemini extracts every transaction into structured data.
- The story has to add up. Documents are cross-checked against each other: running balances must reconcile row by row, rent receipts must match payments leaving an account, transfers out of one account must arrive in another, and salary must arrive regularly. One failed check caps the score at 45. Edit a receipt from 120,000 to 150,000 and it's caught.
- An explainable 0-100 score. Five factors (income stability, payment consistency, savings, balance health, history), every point with its reason. Utility, phone and loan payments count too, so someone who pays no rent can still show a history.
- Six languages. English, French, Kinyarwanda, Spanish, Haitian Creole and Portuguese, including a plain-language summary written by Gemini.
- Share one page. A link or QR code opens a read-only report for any landlord. Landlords see the report, never the files.
- A two-sided rental marketplace. Renters see Miami homes sorted into within reach, almost (with the one tip that would get them there) and not yet, each with a walkthrough video. Landlords list homes with their own minimum score and rent-to-income rules and approve or decline applications in an inbox.
How we built it
- Backend: Python and FastAPI.
- Data: Tiger Data (TimescaleDB). Transactions are a hypertable, and monthly cashflow is a continuous aggregate the score reads from.
- Gemini: structured JSON extraction with a strict schema, multilingual explanations, a fallback chain across models and a local PDF parser so a demo never stalls.
- Frontend: React, TypeScript, Vite and Tailwind, designed with Apple's interface guidelines.
- Auth0: login with JWT-verified API calls.
- Higgsfield: listing photos and walkthrough videos.
- Tests: 42 pytest tests, including the full sample-document story, run by GitHub Actions on every push.
Challenges we ran into
- Gemini quotas and timeouts mid-hackathon. I moved to text-first extraction, a model fallback chain, a result cache and an offline parser, so uploads went from minutes to under a second.
- Building trust without a credit bureau. The answer was making documents vouch for each other instead of trusting any single one.
Accomplishments that we're proud of
- It catches tampering. A faked receipt drops the score from 92 to 45 and names the exact numbers that don't match.
- Kinyarwanda and Haitian Creole sit next to English, so the people this is for can read their own report.
- A complete two-sided product, built solo in one weekend.
What we learned
Verification matters more than scoring. A score is only useful to a landlord if they can see why the evidence agrees.
What's next for Credit Virgin
- Open banking connections for live data.
- More document types (payslips, utility bills).
- Pilots with Miami landlords and university housing offices.
- Credit-builder partnerships, so a strong score can turn into a first US credit line.
Open-source and external code
Built on open-source libraries: FastAPI, asyncpg, pydantic, pypdf, PyJWT, google-genai, React, React Router, Vite, Tailwind CSS, Recharts, qrcode.react, lucide-react and the Auth0 React SDK. Listing photos and walkthrough videos were generated with Higgsfield, and maps come from OpenStreetMap. Everything else (the extraction pipeline, cross-document verification, scoring, marketplace and interface) was written during the hackathon.
Contact
Discord: ophe8724
Built With
- auth0
- fastapi
- gemini
- github-actions
- higgsfield
- postgresql
- pytest
- python
- react
- recharts
- tailwindcss
- tiger-data
- timescaledb
- typescript
- vite