On-device vision inference on commodity hardware, with a central MLOps loop for drift detection and weekly model promotion.
Where it started
Visual quality inspection was manual, inconsistent between shifts, and impossible to audit. A previous vendor's cloud-based system had failed because plant uplinks were unreliable and a round trip cost more time than the line allowed. Lighting and camera positioning differed at every site, so a single model trained centrally degraded badly on rollout.
- 01
Deployed quantised detection models to commodity edge devices, running inference locally so a lost uplink degrades reporting but never stops the line.
- 02
Built a central training loop that ingests sampled frames and operator corrections from every site to produce per-site fine-tunes from a shared base.
- 03
Added drift monitors comparing live prediction distributions against the training baseline, alerting before accuracy visibly degrades.
- 04
Automated weekly champion-challenger promotion with shadow evaluation and one-command rollback per site.
- PyTorch
- ONNX Runtime
- NVIDIA Jetson
- MLflow
- Kubernetes
Where it landed
Twelve plants run the system in production. False negatives — the expensive direction — sit at 0.3%, and per-site models are retrained and promoted weekly without an engineer visiting a factory floor.
A short conversation with an engineer, not a sales qualification call. If we're the wrong people for it, we'll say so and point you somewhere better.