MiMi

Mimi is a team of AI agents for FMCG distributors in Nigeria. A Lagos distributor's vans drop goods at dozens of small shops every day, mostly on credit. The book lives in notebooks and WhatsApp chats, so most fights about money start with "that is not what you delivered" or "I already paid". Millions of naira sit past due, and nobody can prove who owes what. Mimi fixes the paper trail at the source: • After every drop, Mimi sends the shop a receipt on chat and reads the reply in any language the shop uses: English, Pidgin, Yoruba, Igbo or Hausa, typed or as a voice note. A clear yes is kept as the shop's own consent. A dispute goes to the owner in the shop's own words. • When a shop sends a transfer screenshot, Mimi reads every field and code checks it against the open invoice, the account and the date. A fake (like a transfer dated 31 September) goes to the owner, with no blame to the shop. A real one is only "likely paid" until the bank's own credit verifies it. • Mimi estimates stock from deliveries, asks one tap question instead of asking shops to keep a stock book, and drafts the reorder. • For a credit ask, Mimi asks for one shelf photo, SAM 3.1 finds every pack, code matches them to what was delivered, and the owner gets a recommendation that only counts bank-verified payments. The owner sees everything in one console: money at risk, shops, credit limits, and approvals. Mimi proposes, code checks, the owner approves. Mimi never moves money and never changes a limit. Everything in the repo was built inside the hackathon window. The demo runs on a fictional distributor, Coral Lane Trading, with 40 shops and 8 weeks of history. The data is made up. The model calls are real.