Industrial Inspection Vision System

Surface defect detection running on the line itself, trained from a few hundred labelled examples rather than a few hundred thousand.

Client
Placeholder client
My role
ML Engineer
Year
2025
Category
AI Solutions
  • Python
  • PyTorch
  • OpenCV
  • ONNX Runtime
  • NVIDIA Jetson

The constraint was annotation budget. Skilled inspectors were the only people who could label the data, and their time was the scarcest thing in the building.

Approach

Self-supervised pre-training on unlabelled line footage, then fine-tuning on a small labelled set, with active learning to choose which images were worth an inspector's attention next.

Deployment

Exported to ONNX and run on an edge device at the line. No round trip to a server, because the network in that building could not be relied on.

Outcome

Escaped-defect rate roughly halved, with inference under 40 ms on device.

Related work

Clinical Triage Assistant

A triage support tool that ranks incoming cases by urgency and, crucially, abstains when it is not confident enough to be useful.

Read more

Document Intelligence Pipeline

Retrieval and extraction across twenty years of unstructured engineering reports, with citations back to the source page.

Read more