Back to notes
2026-01-20Draft

The vision pipeline on the line: from dataset to Jetson

Training a missing-part detector that has to survive a real, messy dataset and run at line speed on edge hardware.

TODO: scaffold only. Write the post.

The problem on the line

TODO: what missing-part detection has to do, and why late is expensive.

The dataset was the real adversary

TODO: why the class taxonomy had to be redesigned against real data, not a clean benchmark.

From training to TensorRT on a Jetson Nano

TODO: the export path (ONNX → TensorRT), what it bought in latency, what it cost in portability.

Closing the loop with traceability

TODO: how detections feed back into the traceability system.