Real-time object detection across three production lines
96.4%
Detection accuracy in production
<3%
False-positive rate
70%
Less manual inspection time
12,000+
Images processed daily, no downtime
The challenge
Visual inspection ran manually across three production lines. Automating it was only worth doing if the model cleared 95% detection accuracy while keeping false positives under 3%, because below that bar the team would spend more time reviewing the system than it saved.
Our solution
We trained a YOLOv8-based detection model on their production imagery and exposed it through a REST API so it could sit inside the existing inspection workflow. The model runs on edge devices at the lines rather than in the cloud, which keeps the decision local and fast. Production monitoring tracks throughput and detection behaviour so drift surfaces before it reaches output.
Tech stack
“Overall, we were very satisfied with their work.”
Managing Director
Lunara co sp. z o.o. (verified Clutch review)
Need similar results?
Let's discuss your project.