Conforme

Every morning my mother signs a pile of supplier delivery notes at our family restaurant in Gijón, Spain, while the driver waits. She doesn't have time to check the sums, and once a note's signed, a wrong line is hard to claim back. I built Conforme so she can check them before she signs. You take one photo of the counter. SAM 3.1 finds each note and straightens it. Muse Spark 1.3 reads every line with its box on the paper, then reads the money again on its own, without seeing the first reading. My code checks every euro: quantity times price, VAT, totals, tax IDs, prices against past deliveries, the order and duplicates. When something doesn't add up, an agent goes back to the paper. Muse Spark with tools decides what to look at. It re-reads a cell or a row zoomed in, pulls the price history or the order, and closes the case citing numbered evidence. My code checks every piece of evidence it cites, and if a second look disagrees, the claim's held back and the note goes to her. Before it asks a supplier for a cent, the agent confirms on the paper each number the claim rests on. What she gets is the line to write on the note before she signs, and a message for the supplier with the proof cut from their own paper. Nothing's sent without her. You can also just ask, out loud: "how much have I overpaid this month?" Muse Voice Transcribe and Spark turn that into an intent, my code works out the figure, and the lines behind it light up on each note. I measured it on 33 real delivery notes from our kitchen. 20 of them were scanned sideways, some are faded dot-matrix prints, a few have two VAT rates on one page. In two full runs it read 98.7% of the lines to the cent (152 of 154), and it never stamped ALL GOOD on a note it had misread. It costs me about $0.036 per note. The real notes stay on my laptop; only the numbers are public. There's a private mode too. One setting sends every Spark call to Muse Spark Open, Meta's open-weight model, on Baseten, so a restaurant's prices don't have to go to a closed API. On the 16 fictional sample notes it read 38 of 38 lines in its first run, and you can watch the counter sample read by it on the live site, next to Muse Spark 1.3.