For two years, our defect-detection model lived in a data center three states away. Images left the cell, a batch job scored them, and a result came back forty to ninety seconds later. That was fine when the line ran at a pace that could absorb the delay. It stopped being fine the day we added a 0.6 second cycle time to the Austin cell.
The move to the edge
We re-trained the same model to run on an industrial PC bolted to the cell controller, four feet from the camera. Inference now completes in under 18 milliseconds. No network hop, no queue, no batch window.
The hard part was not the model. It was proving that a model running on a fanless box with no GPU would match the accuracy of the cloud version. We ran both in parallel for six weeks before we trusted the edge result on its own.
What we measured
- False-reject rate on the edge model: 1.2 percent, against 1.4 percent in the cloud version.
- Median latency: 18 ms edge, 61,000 ms cloud (batch window included).
- Units reworked due to late detection, Austin cell, trailing 90 days before the switch: 340. Trailing 90 days after: 9.
What we would do differently
Start the parallel run earlier. We spent three of the six weeks arguing about whether the comparison was fair before we started collecting the data that settled it.
The edge model is now standard on every new cell we commission. The cloud model still runs, but only for retraining and for cells where the camera resolution does not justify local compute yet.