# Grounded

- **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:** [Tianqi Yang](https://cerebralvalley.ai/u/tianqiy)
- **GitHub:** https://github.com/tianqiyang/grounded-coach
- **Demo video:** https://youtu.be/s45n-gnx6Xo
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1417

AI interview coaches make your answers sound better by making things up, and that inflated number is the follow-up question that sinks you. In our eval, a naive "make it impressive" prompt put unsupported claims in 18 of 20 rewrites; even "use only these facts" still did in 7 of 20.

Grounded is an interview coach that never invents your experience. Muse Spark extracts facts from your résumé, each with a verbatim source quote, and nothing counts until you confirm it. Paste a job description and it predicts 8 likely questions, matches each to your best story, and names your gaps. Answer by voice or typing and you get scores, a claim-by-claim breakdown, and a warning for any number that isn't in your record. Then a skeptical follow-up drill rewards "my record doesn't cover that" over filling the gap, and you leave with a one-page prep brief.

Every sentence of the rewrite passes two checks: local code verifies it cites a confirmed fact, quotes it word for word, and uses only that fact's numbers; then a narrow Muse Spark entailment pass removes anything that says more than its fact. With both checks, unsupported rewrites dropped to 0 of 20. Our first run, with the verifier alone, didn't beat the instructed prompt, and both runs are in the repo. All data is fictional.

## More from Meta Global AI Developer Hackathon

- [FlowForge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1414)
- [Punta de Lanza — PROOF](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1415)
- [IdeaNirvana](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1416)
- [GLTrax](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1418)
- [OfferFlow AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1419)
- [AccessiWebEngine](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1420)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1417. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt