readykit-edge

Emergency kits fail quietly. A trauma kit can look complete and still be missing its shears, hold one tourniquet where it needs two, or carry sealed gauze that expired last month. These kits are checked by eye, in a hurry, by people who have to trust the result. Automating that check with the obvious approach, asking an AI "is anything missing?" and searching the reply for "missing" or "no", fails in a dangerous way. A reply of "Absent from the tray: Trauma Shears" contains neither word, so that design unlocks a kit with no shears in it. We wanted a system that refuses to open unless it's sure. ReadyKit Edge is an air-gapped kit inspector with a physical lock. A camera watches the kit, a vision-language model running on the device checks every required item against a manifest, and the result drives a latch. Every inspection ends in one of three verdicts: PASS, FAIL or INDETERMINATE. Only PASS releases the latch. It reads use-by dates and counts quantities, not just presence. If the model can't tell, answers vaguely, or crashes, the latch stays locked. Every inspection is written to a tamper-evident, hash-chained audit log. An operator console shows the live camera, the exact frame judged, and what the model recognised. Nothing leaves the device: no cloud, no network calls. How we built it Qualcomm Snapdragon X Elite runs Qwen3-VL-4B through GenieX on the Hexagon NPU. Ollama with qwen2.5vl covers other machines. Arduino UNO Q firmware drives the latch. The loop never blocks, and a watchdog locks the latch if the host goes silent. A custom serial protocol carries every command with a CRC-8 checksum, a sequence number and an acknowledgement. Python for the decision engine, FastAPI for the console, OpenCV and the browser camera for capture. On real hardware the model invented a complete trauma kit, with all seven items found at over 95% confidence.