# QuantProof

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Joyeb Kashyeb](https://cerebralvalley.ai/u/ihs)
- **GitHub:** https://github.com/Kashyeb57/quantproof
- **Demo video:** https://youtu.be/2x6z8SFrNXs
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/702

Live app: https://kashyeb57-quantproof-app-snew9v.streamlit.app/

Banks sell structured notes to ordinary investors. The payoff rules are buried in long SEC filings, and the bank's own estimated value is usually below the price paid. QuantProof uses Meta's models to read those filings, and its own code to check the evidence and compute an estimated value range next to the bank's estimate.

The frontier is verified AI-written financial code. The AI reads and writes; code calculates and checks.
- Every extracted term carries a cited passage, and code verifies it word for word against the filing.
- A second, independent extraction by Meta's open-weight model compares the terms field by field.
- When a note has a feature no template covers, a Meta model writes the payoff code. It runs in a sandbox (subprocess, timeout, no network, import allow-list) and is trusted only after it passes numerical checks, including the filing's own worked payment examples to the cent. Failures send the exact counterexample back for repair.
- The AI never states a number: every figure comes from the engine or a cited filing value, and a provenance check enforces this.

Results from the saved offline studies:
- 44 recent filings from 11 issuers read; 42 extracted.
- 95.2% of cited fields verified against the filing text.
- The bank's estimate fell inside our range on 11 of 13 priced notes.
- 24 of 31 AI-written payoffs passed, 5 of them after repair.
- Coverage rose from 11 to 26 of 42 notes.
- Total model cost: $12.07.

Risk-neutral values and real-world loss odds are computed and labelled separately. Results are ranges under stated assumptions. This is risk analysis, not investment advice.

## More from Meta Global AI Developer Hackathon

- [Codicent](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/699)
- [1556836796459675](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/700)
- [Habitat Watch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/701)
- [DriveMatch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/703)
- [Theotech](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/704)
- [Intenzo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/705)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/702. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt