Marea: sales copilot for small-scale fish sellers

THE PROBLEM Small seafood sellers in Honduras sell on WhatsApp and Facebook Marketplace. Fish spoils in 1-3 days and every buyer haggles. The seller answers by hand and guesses: they refuse offers on fish that will spoil anyway, or accept low offers on fish that would sell at full price. WHAT MAREA DOES 1. Reads a buyer message ("te doy 700 lps x las 10 lbs, me las lleva mañana a Sambo Creek") and shows DECIDE TÚ when a deal needs the seller: offer below the floor, low stock, freshness, zone, credit. 2. Sell now or wait: compares accept, counter and wait in lempiras. Demo: accepting L 70/lb brings L 700 more than waiting, because 23 lb would spoil. 3. Several buyers, one lot: ranks who to serve first. Demo: L 3,110 today vs L 1,590 in arrival order. 4. Dashboard and today's actions: freshness, stock at risk, markdowns, flash offers and a daily post. Nothing is sent automatically. WHERE AMD RUNS Gemma-4-31B-IT on AMD GPUs through AMD Token Factory. Gemma reads the messages the rules cannot (slang, typos, English fish names) and writes the replies. Code keeps only numbers written in the message and checks every price the model writes. On 10 hard messages: rules alone 0/10, rules + Gemma 10/10. All math runs in pandas; the model never sets a price. TESTS AND LIMITS 30/30 synthetic Honduran WhatsApp messages read by the rules, plus unit tests. All data is synthetic. No WhatsApp integration yet: the seller copies and pastes.