# umwayi

- **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:** [Miguel Shinyenyi](https://cerebralvalley.ai/u/Amwayi)
- **GitHub:** https://github.com/Miguel-Shinyenyi/umwayi-rm
- **Demo video:** https://www.youtube.com/watch?v=XLmpeBoWaes
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3890

Umwayi-RM is an open source secretary for savings groups (chamas). Treasurers keep records on paper and in chat messages, where names collide, months go missing and totals disagree. After a meeting, the treasurer sends a voice note or a photo of the paper page. The agent returns proposals. It never writes the ledger.
 
How it works
- A voice note is transcribed by Muse Voice. Muse Spark Open reads the text and proposes a ledger entry. Every proposal quotes the exact words it came from.
- When a note is ambiguous, for example two members named Brian, the agent asks one question instead of guessing. The treasurer answers and a proposal follows.
- A photo of a handwritten page is read by Muse Spark Open. Plain code compares the rows with the stored ledger and lists three kinds of difference: on the page only, in the ledger only, and different amounts. They are questions for the treasurer, not accusations.
- Nothing changes until the treasurer approves. Approval is idempotent, and a second model call checks that the quoted words support the entry. The ledger is append only and owned by code.
- A Today list shows who still owes. The agent can draft reminders as proposals through its tools. Nothing is sent.
- A Lab page shows the eval results and the model id of every run.
 
Built during the hackathon. Live demo: umwayi-rm.vercel.app (synthetic data only). Code: github.com/Miguel-Shinyenyi/umwayi-rm (MIT).
 
What I measured (synthetic data, frozen test split, 3 seeds, 40 notes each)
- Clear notes handled correctly, out of 72: curated prompt pack 67, same-length placebo 55, generic prompt 24, bare prompt 23. About three quarters of the gain comes from prompt length, not content.
- Ambiguous notes, wrong guesses out of 48: 0 for curated, placebo and generic; 7 for bare. My claim that the curated pack adds restraint is not supported.
- The archive memory test did not support its claim either.
 
Not proven: that treasurers prefer it to notebooks or to existing apps, that it reduces errors in real groups, and security or Data Protection Act compliance. I have not interviewed treasurers, so the problem is a hypothesis.
 
Ask: one pilot group for a month, and feedback from treasurers.

## More from Meta Global AI Developer Hackathon

- [Byte me](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3887)
- [Causal Delta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3888)
- [AMONS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3889)
- [ScottB](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3891)
- [ScreenTutor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3892)
- [Buzzed Cut](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3893)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3890. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt