Order Desk

Order Desk is an AI desk that learns each distributor's language. Muse Spark writes its own matching rules, a second Spark call attacks them with look-alike products, and only rules with zero wrong matches go live. It turns WhatsApp orders and supplier invoices (Darja voice notes, photos of lists, French and Arabic mixed) into verified transactions in the distributor's own ERP. The problem: across Algeria and North Africa, wholesale runs on WhatsApp. At every distributor, a clerk retypes each order and each supplier invoice into the ERP, guessing which of 1,000 products "ifri 1.5" means. Apps that ask shops to change their habits don't get used: I built one, and nobody used it. So Order Desk goes where the orders already are. What it does: - Shops keep ordering on WhatsApp or Telegram, by text, voice note or photo. WhatsApp runs on Meta's Cloud API (signed webhook, reply buttons); the public demo uses Telegram so anyone can try it without a WhatsApp Business number. - Every line climbs a ladder, cheapest first: product codes, trusted wordings, Spark-written rules, a shortlist built by code, one Muse Spark pick from that shortlist, then a person. Each line is checked against the ERP's prices and stock. - Not sure? It asks the shop one question with buttons instead of guessing. Confirming twice never books twice. - Supplier invoices: Muse Spark and Muse Wonder read each photo or PDF independently. A line they disagree on goes to a person; the rest becomes a goods receipt. - Safe by design: the webhook is the only public route and every call must carry Meta's signature; keys never reach logs; ERP passwords are not kept. - One adapter API, five open-source ERPs (Odoo, ERPNext, Dolibarr, InvenTree, Tryton), each passing the same live conformance test, with nothing installed in the ERP. Measured on a benchmark I built from scratch (4 real catalogues, 760 supplier invoices, 200 shop orders, all with answer keys). Every number reruns from scripts in bench/ (bench/check.py): - Invoices: 93% of lines booked automatically on 120 locked invoices the desk never learned from, 3 wrong in 2,169 lines, 58% with no model call at all, 4-8x cheaper than Muse Spark alone with the same 3 wrong. - Shop orders (French, Darja, Arabic-script quantities, screenshots, voice notes): 76% of 1,272 lines booked right, most of the rest asked, only 3 wrong products. - 23 real, crumpled delivery notes: Muse Wonder read 98% of 383 lines right, Muse Spark 93%; where both agree, 1 line wrong in 317. Built solo by me, Nadir, from Algeria, during the event: about 22,000 lines of Go, 79 tests, a hosted live demo. Live demo (up until Nov 8): - Panel, read-only: https://orderdesk-demo.elixir-dz.com - Telegram bot: https://t.me/elixirOrderDeskbot code META26 - Odoo, read-only: https://odoo-demo.elixir-dz.com judge / Meta2026! - Locally, no key: go run ./cmd/orderdesk -demo