Zambrite

Mesa Muse gives a market trader a shop on WhatsApp built from one voice note and one photo. The problem: most small traders in Zambia sell from a market stall, a table by the road or a corner shop, and run the business from WhatsApp. Customers message, the trader answers when they can, and sales are lost while they serve someone in person. Putting stock online means typing names and prices into an app, in English, on a small phone. Most traders never do it. How it works, all inside WhatsApp (official Cloud API only): 1. The owner sends a voice note in Bemba, Nyanja or English describing the shop. It becomes a shop profile (name, what they sell, location, hours, payments, delivery) with one-tap confirm buttons. 2. The owner sends one photo of the table or shelf. Muse Spark names each product, SAM 3.1 cuts every one out along its pixel mask, and Muse Spark prices the numbered cut-outs in kwacha. 3. The owner corrects by voice: "number 2 is K90, take out the soap". The catalogue updates. 4. One tap opens the shop: the owner gets a QR code for the stall and a wa.me link. 5. Customers type, send a voice note, or send a photo ("do you have this?") in their own language. An assistant shows product cut-out cards, builds the cart and places the order. 6. Every order reaches the owner with Accept / Can't-do-it buttons, and the customer is told either way. If a customer asks something the shop facts don't cover, the owner gets an "Answer them" button and their voice note goes straight to that customer. A live control room shows every WhatsApp message and every Meta model call (task, latency, output) as it happens. Built for production: deployed at https://muse.zambrite.com on our AWS infrastructure (Docker, Terraform, CodeBuild/CodeDeploy). HMAC-verified webhooks, de-duplication of WhatsApp retries, per-sender queues, a control room that refuses to start without a password, and 17 automated tests that run a full owner-and-customer journey through the real webhook against fake Graph and Meta API servers. Status, stated plainly: our dev.meta.ai account was restricted when the temporary card hold failed, so we could not call the live Meta Model API before the deadline (the $150 credits are on the account). The demo video runs the real Mesa Muse code with scripted model and WhatsApp responses and says so on screen. The model calls follow Meta's cookbook request shapes. The only missing piece is the API key; then `npm run doctor` checks each model and it runs live. What's next: mobile-money payment (MTN MoMo, Airtel Money) and rollout to the existing business customers of Mesa, Zambrite's WhatsApp business product.