# TraceForm

- **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:** [Nana Kojo Atta-Benyah](https://cerebralvalley.ai/u/htaik)
- **GitHub:** https://github.com/Nana-Kojo801/traceform
- **Demo video:** https://youtu.be/ge49UzfQIRo?si=hmKPDeHHqx3GhKbv
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3454

TraceForm fills in forms from messy documents with Meta's Muse Spark model, and shows proof for every answer.

THE PROBLEM
Small organizations (nonprofits, community groups, small vendors) spend hours copying facts from emails, letters, meeting minutes and scans into intake and grant forms. AI can fill a form in seconds, but you can't tell which answers are real, which were guessed, and which quietly picked one of two contradicting numbers. Checking the result means re-reading every source, which removes the time saved.

HOW IT WORKS
You upload up to five documents (TXT, DOCX, PDF, or photos of pages) and define a form. Muse Spark answers every field with a verbatim quote from a specific document. Then plain code, not the model, searches the cited document for that quote. A quote that isn't found is never shown as verified. Each failed field gets one repair pass in which the model must re-cite word for word or give up. When documents disagree (for example $48,000 in an application vs $54,000 in a later email), TraceForm shows both sides with their sentences and a person chooses. If no document states a fact, the field stays blank. A person approves every answer, then exports CSV or JSON of the reviewed values. Photos and scanned PDFs are read by Muse Spark vision and labeled "model transcription".

WHAT MAKES IT DIFFERENT
Verification is deterministic and separate from generation. Contradictions are surfaced, never silently resolved. Undocumented facts are never guessed. And the UI keeps "the quote exists" (checked by code) apart from "the quote answers the question" (a human decides).

MEASURED RESULTS (EVALUATION.md in the repo)
On a fictional grant-intake document set with planted facts, 5 live runs gave 45/45 correct fields, 0 invented values, the planted conflict flagged 5/5 and the undocumented field left blank 5/5. In separate live probes, 3/3 deliberately fabricated citations were caught by the verifier; the model re-cited the 2 real facts verbatim and withdrew the fake one. One sample run in the demo is precomputed fixture data and is labeled as such. This is a small synthetic test, not a benchmark.

STACK
React, Vite, TypeScript, Tailwind, Convex (database and server actions), Zod. Live demo: https://traceform-five.vercel.app/
Demo data is fictional. TraceForm is a hackathon demonstration, not certified for any regulated use.

## More from Meta Global AI Developer Hackathon

- [Nanjo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3451)
- [Autumn's Second Bloom](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3452)
- [AVO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3453)
- [ASCII](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3455)
- [Amos Kanyanta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3456)
- [Nearhand](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3457)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3454. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt