# Duka Books

- **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:** [FONDO ADAM](https://cerebralvalley.ai/u/fondoadam)
- **GitHub:** https://github.com/fondoadam/duka-books
- **Demo video:** https://youtu.be/oRIlghZCtjc
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/587

Duka Books gives Kenya's corner shops real books without an app, a form or an accountant.

The problem: Kenya has about 7.4 million micro, small and medium establishments (KNBS 2016), and the neighbourhood duka is the most common. Most keep records in an exercise book and money in M-Pesa. A lender can't read either, so credit often means Fuliza overdrafts and other costly short-term loans. Bookkeeping apps ask owners to change how they work, in a language that often isn't theirs.

What it does: owners send what they already have to a WhatsApp number, in Swahili.
- Ledger photos: Muse Glimmer 30B reads handwritten pages and checks them against the written total.
- Receipt piles: SAM 3 finds each receipt in one photo and numbers it, and each crop is read on its own.
- M-Pesa and Fuliza SMS: exact parsers, with no model. A Fuliza draw is booked as a loan, not income, and a transaction code can't be booked twice.
- Confirm or fix: nothing reaches the books without a tap. "Si sahihi" opens a line-by-line fix, and each fix becomes an example for that shop.
- Questions: "faida wiki hii?" is answered by Llama 3.1 8B calling tools over the journal. A figure guard rejects any number no tool returned.
- Reports: an evening summary and a bilingual monthly profit-and-loss PDF a SACCO or bank can read.

What's live: the WhatsApp bot on a Meta test number. A real phone sent an M-Pesa SMS and a handwritten ledger photo through to the journal on Nov 1, then asked for its profit and P&L PDF. Voice notes still use a stub transcript. All figures are test data; it has not yet been used by a real shop.

All code was written during the event, from an empty repo (first commit Oct 27, 2026).

## More from Meta Global AI Developer Hackathon

- [Convoplex Retention Agent](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/584)
- [FlexVertex](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/585)
- [Course Factory](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/586)
- [StationSync LLC](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/588)
- [Suite Fleet](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/589)
- [Blackout](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/590)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/587. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt