# LedgerLens

- **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:** [Aaron Malunga](https://cerebralvalley.ai/u/Oraan)
- **GitHub:** https://github.com/aaronmalunga/LedgerLens
- **Demo video:** https://youtu.be/nxw78IfoQV0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3088

Spaza shops in South Africa and tuntemba (roadside kiosks) in Zambia already keep records: receipt books, till slips, notebooks, mobile-money messages and voice notes. Lenders cannot use them, so traders with real trading histories stay credit-invisible. 

The scale is large: 76% of Zambia's workers are informally employed (ZamStats Labour Force Survey 2022) and 5.7 million South Africans are in informal employment (Stats SA, Q4 2025); missing financial records are a reported reason Zambian small businesses are refused loans (2022 Bank of Zambia MSME Finance Survey, as reported by BongoHive, 2025). LedgerLens turns those records, as they are, into a ledger the trader confirms and a credit file a microfinance officer can read. 
Muse Spark 1.3 reads photos of printed and handwritten records into a fixed schema; Muse Voice Transcribe turns voice notes into entries. Code, not the model, does the arithmetic and assigns an evidence tier to every line (third-party record with an ID, receipt without one, or self-reported), and nothing is saved until the trader confirms. The credit file shows monthly money in and out, how much income is backed by third-party evidence, and the source of every line; it never scores the trader.
Measured on the developer's own real South African and Zambian receipts (black-boxed): on a sealed hold-out scored once, totals correct 6 of 7; on the practice set, whole record correct 21 of 24 and totals 23 of 24, at about $0.01 and 11-16 seconds per record.

 A person typing the same records took a median of 29 seconds each at the same accuracy. Voice notes, including one in code-switched "town Nyanja", were read correctly 3 of 3; 6 of 6 prompt-injection attacks were resisted. 
p95 latency 27.5 s; 0 errors on the hold-out. Known limit, disclosed: a photo that is not a financial record can still produce a made-up amount; the warnings show and the trader rejects it. Live web demo (WhatsApp-first design; the web chat is the planned fallback because Meta's test account blocked outgoing WhatsApp messages).

## More from Meta Global AI Developer Hackathon

- [VaultLine](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3085)
- [Replace Consulting](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3086)
- [Seventy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3087)
- [Balance Dex](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3089)
- [MiMi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3090)
- [See The Germs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3091)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3088. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt