Emanuele Menon@emnlmn47846

Wholesale distributors still take orders the way their customers talk: a WhatsApp message at 6 a.m., a voice note from the van, a photo of a handwritten list. Someone at the order desk retypes each one, guessing what "the usual" means and whether the price is the one agreed. CatOrder does the reading. Gemma 4 31B, served on AMD Instinct GPUs through AMD Token Factory, extracts every line from text, voice-note transcripts and photos, matches it to the catalogue, and resolves "the usual" or "same as last week" from the customer's order history. Each line is checked against the customer's agreed terms, so a wrong price, an odd quantity or an ambiguous product gets flagged instead of slipping through. The order desk gets a draft order where every line points back to the exact words it came from. A person confirms; nothing is sent on its own. When something is unclear, the model drafts a WhatsApp reply for the operator to edit and send. Real messages arrive live through the Twilio WhatsApp sandbox. On 20 synthetic messages (WhatsApp, photo, voice, email), Gemma 4 31B on Token Factory got 70 of 70 lines and 20 of 20 orders right, and found all 8 expected flags with one false flag. The whole run took 84k input and 6.6k output tokens, about 21 seconds per message. Built in TypeScript with Hono and SQLite, with a plain frontend and no framework. All data is synthetic. MIT licensed.

Food and beverage distributors still take orders the way their customers talk: a WhatsApp at eleven at night, a voice note from the car, a photo of a handwritten list. Someone at the order desk retypes each one, guessing what "the usual" means and whether the price is the one agreed. Nobody notices the bar that has gone quiet until it has moved to a competitor. CatOrder does the reading. Gemma 4 31B, served on AMD Instinct GPUs through AMD Token Factory, extracts every line from text, photos and voice-note transcripts. Code then matches each line to the catalogue, resolves "the usual" from the order history and checks the agreed terms, flagging wrong prices, odd quantities and ambiguous products. Every line points back to the exact words it came from. A person confirms, and Gemma drafts the WhatsApp reply. CatOrder also looks ahead. catorder-nbr, a small PyTorch model trained with ROCm on an AMD Radeon Pro W7900D, predicts each customer's next order date and basket: who has gone quiet, what they forgot, what to buy next week. Results, on synthetic data: on 237 WhatsApp and email messages never used to tune the prompt, 232 orders entirely right (98%) and 99.7% of lines right. Every miss was reviewed by hand (217 of 237 before that review); the rules are in the repo. On 135 held-out customers the forecast beats the strongest simple baseline on basket F1 (0.823 vs 0.803), ties it on the date and trails it on churn. Live WhatsApp through Twilio; TypeScript, Hono and SQLite; MIT licensed. Next: connectors to the back-office systems small distributors already use, now cheap to build with AI-assisted coding. Built by Emanuele Menon.

Six Matches. One leaderboard. Build on AMD, earn points and compete for prizes.

🌎 Hybrid Hackathon | 💻 Online Build: 12-17 October 2026 | On-site phase: 17-18 October 2026 To participate in the hackathon, simply click Sign up with AMD. If you are not already a member of the AMD AI Developer Program (ADP), you will need to create an account. The AMD Developer Hackathon: ACT III is a hands-on event for developers, founders, engineers and builders who want to push what’s possible with AI on real infrastructure. 📍 If approved for the on-site experience, you'll also have the opportunity to showcase your project live. Travel and accommodation will not be covered.