Chama Books

A UGREEN AI NAS Skill that closes the month for a savings group: it checks M-Pesa screenshots and the meeting recording against the real statement, and flags payments that never arrived.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Chama Books

Inspiration

Almost everyone in Kenya belongs to a chama, a savings group where members pay in every month. Members pay by M-Pesa and post screenshots in the WhatsApp group, and the treasurer is a volunteer. At month end that volunteer scrolls through hundreds of messages to match screenshots against the M-Pesa statement. One wrong figure and the next meeting turns into an argument. Fake M-Pesa confirmation messages are common, so a screenshot on its own proves nothing.

What it does

Chama Books is a Skill for UGREEN AI NAS. The treasurer keeps the statement, screenshots and meeting recording in the group's NAS folder, then says "Close October for Umoja Chama". The Skill:

  • finds the month's files
  • reads the statement and register, and OCRs every payment screenshot
  • transcribes the meeting, picking out who said they paid, loans and fines
  • reconciles everything against the statement, which is the only source of truth
  • lists exceptions: screenshots whose code isn't on the statement, money from unknown numbers, non-payers, double claims
  • asks the treasurer about each odd case, then, after approval, saves the treasurer report, member statements and draft minutes, and tags unresolved screenshots

How we built it

The Skill is an orchestrator written with UGREEN's uliya-skill-kit. Nine steps are delegated to file-agent, document-agent, photo-agent and knowledge-agent. Each step only uses results the previous expert actually returned. Reads need no approval; anything that writes or tags waits for the treasurer. It passes the kit's static check.

Challenges

M-Pesa statements and screenshots vary a lot, and meetings switch between Swahili and English. Both need testing on a real NAS. The design rule we kept coming back to: never guess a transaction code, and never mark someone as paid on a screenshot alone.

What's next

Build a synthetic test chama (12 members, statement PDF, screenshots, a short recording) and run the core and edge cases on the NAS if we reach the finals.

Disclosure: Claude (Anthropic) through Claude Code helped research the kit and draft the Skill and text; I reviewed and edited it. M-Pesa is a Safaricom service; this project is not affiliated with Safaricom. All test data is synthetic.

Built With

Share this project:

Updates

Submission history