Utpal DasML Defect Detection for CT Scans
Training neural networks for industrial inspection in 2017 — years before AI was fashionable.
Client
My role
Project lead — requirements through delivery
Team
In collaboration with Fraunhofer
Year
2017 – 2018
Stack
Categories
The problem
Precision pipes for semiconductor manufacturing had to be checked for micro-defects. A trained human spent a week per batch squinting at computed-tomography scans — slow, expensive, and inconsistent in exactly the way a person gets tired.
What I built
- Trained a Caffe model on labelled defects, classifying by probability, type, size and rest-wall thickness.
- Built volumetric scan import so CT data went straight into the pipeline.
- Designed a trainable master/user workflow so engineers could teach the model new defect classes as they appeared in production.
- Led the project from requirement definition through delivery, with Fraunhofer researchers advocating for neural nets harder than I did.
The outcome
It worked — despite near-universal scepticism at the time. My first AI project, eight years before the rest of the industry arrived.
The story
My first AI project. 2017. Pre-Transformer, pre-hype. In collaboration with Fraunhofer, whose researchers advocated for neural nets harder than I did.
The problem: precision pipes for semiconductor manufacturing had to be inspected for micro-defects. A trained human spent a week per batch, squinting at CT scans. We trained a Caffe model on labelled defects — probability, type, size, rest-wall thickness — with a master/user workflow so an engineer could teach the model new defect classes as they appeared in production.
Every time I told someone what I was working on, they thought it wouldn't work. It worked. That's when the 2008 daydream came back to me — a world where you type in English and machines write code. Fifteen years early. But the direction was right.