Ironclad Fabrication
A computer vision system that inspects welds on the assembly line in real time, catching defects before a part reaches final assembly.
Industry
Manufacturing
Timeline
15 weeks
Team
2 ML engineers, 1 computer vision engineer
Category
AI Development
Where things stood
Ironclad's quality team was spot-checking welds by hand, sampling roughly one in ten parts, which meant defective welds sometimes reached final assembly before anyone caught them.
How we got there
Line-side camera capture
Installed cameras at each welding station capturing every part instead of a manual sample, feeding images straight into the inspection pipeline.
Defect classification model
Trained a computer vision model on a year of labeled weld images to flag porosity, undercut, and misalignment in real time.
Line integration with operator alerts
Wired the model's output into the line's existing alert system so an operator gets flagged immediately, not after a part moves three stations down.
What changed
100%
Of welds inspected, up from ~10% manual sampling
<1s
Inspection time per part
~40%
Reduction in defective parts reaching final assembly
Tech stack
Python
The language for the inspection pipeline and model training.
PyTorch
Trains the defect-classification model on labeled weld images.
OpenCV
Handles image preprocessing and camera capture at each station.
AWS SageMaker
Hosts the model and serves inspection results in real time.
PostgreSQL
Logs every inspection result for quality audit and model retraining.
Docker
Packages the inspection service identically across every line-side camera station.
