Your new AI inspection system catches every defect at 97% accuracy. In the lab. On the press, it rejects 14% of good parts. A stamping shop mounts two cameras on a 200-ton press running 40 strokes a minute. In testing, the model nails burrs, short fills, and slug marks. In production, false rejects jump to 14%. Why? The lab used controlled ring lighting. On the floor, sheet metal reflects the overhead bay lights differently at every stroke angle. Lubricant pools differently on a warm die than a cold one. The first 50 parts of a shift don't look like parts at thermal equilibrium. The fix is not a smarter model. It's polarized lighting to kill the glare, a thermal camera to correlate surface appearance with die temperature, and training images from cold-start, mid-run, and end-of-run conditions. Then the real work starts: getting the reject signal to the PLC inside the 750ms stroke window, tagging each part in the MES, routing defect images to the right engineer. That integration is roughly 60% of the project timeline. Model training is 15%. The hardware is a purchase order. It's why 84% of these integration projects fail or stall — legacy systems that can't talk to each other. The Jetson runs inference in 12ms. That was never the hard part. Out-of-the-box AOI throws 5-15% false rejects. Well-tuned systems land under 2%. If you've put vision AI on a line, what broke first in production — the lighting, the data plumbing, or the OT integration? #EdgeAI #SmartManufacturing