A supplier sends a new price list. Brake pads went from 9.60 to 5.10. Good news? No: the pack went from four to two, so each pad now costs 6% more. A small e-commerce gets dozens of these lists, and nobody checks every line. Margin Guardian catches every hidden price increase. It reads the price lists (PDF, Excel, or a phone photo of a printed one) in the supplier's own language and currency, compares them with the previous ones, links each line to the catalog and computes what changed: unit cost in euros, margin before and after, cost per month. On the demo data it finds 19 changes in 3 price lists, 7 of them hidden, worth about 9,150 euros a year, including a surcharge written only in a footnote. It doesn't stop at a warning. For each increase it puts three moves side by side, with numbers: absorb it, reprice (capped by the cheapest competitor, with lost sales estimated from 24 months of history) or negotiate. For the negotiation it writes the email to the supplier. Nothing is sent: a person reads it and presses the button. The AI reads and writes, the code calculates, the person signs. Margins and euros come from Python, not from the model. A reviewer written in code checks that every cost comes out of a price written in its source row, and blocks it otherwise. Click any number and the document opens with that row highlighted. A photo is read twice, and a row counts only if both readings agree. Built on Gemma 4 31B on AMD Token Factory, the Evolus extractor for PDF price lists and the Evolus email tool (Evolus workspace: Voren). If Gemma does not answer, Qwen on the same AMD servers reads the photo, and the app says which service failed. Competing in the Reinvent Commerce track, for the Evolus Business Agents prize and for the Google Gemma prize.
Category tags: