.png&w=256&q=75)
🤓 Latest Submissions
Foreman: Find it. Explain it. Fix it.
On a metal-forming line, a worn die or failing lubrication costs twice: scrap made before anyone notices, then downtime while a technician works out what the symptom means. Cameras can say a part is bad, but not why. That answer lives in a thick manual and in senior technicians' heads. Foreman takes a plant from "this part looks wrong" to an approved work order: 1. Detect: PatchCore, trained only on good parts, scores every part and outlines the defect on a heatmap. 2. Fuse: press sensors are checked against the machine's baseline with plain statistics. 3. Retrieve: the manual sections that match the defect and sensor pattern. 4. Reason: Gemma 4 vision reads the photo, heatmap, sensors and manual, and writes the cause with a confidence, the evidence, what it ruled out, repair steps and an operator message. 5. Verify: every manual citation is checked against what the model was shown. 6. Approve: a supervisor signs off, and the sensors return to normal. Sensors usually drift before the camera sees bad parts, so Foreman raises an early warning first and escalates the same incident when the camera confirms it. Why AMD: everything runs on one AMD Ryzen 7 7435HS CPU, with no GPU and no cloud. Gemma 4 E2B runs on llama.cpp and PatchCore on PyTorch, both CPU-only. Camera images never leave the building, which many manufacturers require. A ROCm + vLLM setup for Gemma 4 31B on AMD Instinct MI300X is included for scale; it is built but not yet run. Measured on the Ryzen: 0.996 image AUROC on MVTec AD metal_nut, 0% false alarms, about 184 ms per part. In the live eval, 12 of 12 faults got the right defect, camera/sensor agreement and manual citation. Median diagnosis time is 2.8 minutes, with 34 automated tests.
18 Oct 2026
