Computer vision is transforming quality control on manufacturing lines — but the gap between a proof-of-concept and a reliable production system is larger than most teams expect. This is a first-person account of what it actually takes, based on our deployment at ManuTech Albania.
The Setup
The production line runs at 120 units per minute. Each unit needs to be inspected for 23 defect types — scratches, warps, dimension deviations, and surface blemishes — that were previously caught (or missed) by human inspectors at end-of-line. Miss rate with manual inspection: 1.8%.
Data Collection is the Real Challenge
We spent the first three weeks on data collection, not model development. Getting 50,000 high-quality, annotated images across all defect types, lighting conditions, and product variants is genuinely difficult on a live production line. We built a custom image capture rig and worked with quality engineers to develop annotation guidelines that matched their actual defect definitions.
The annotation process revealed disagreements among the human inspectors themselves — which is itself a valuable finding. We used this to produce clear, written defect taxonomy documentation that the company had never had before.
Edge Deployment Realities
Running inference on NVIDIA Jetson at line speed sounds straightforward until you encounter vibration interference, inconsistent lighting, and the operational constraint that the system cannot slow the line. We went through four hardware iterations before we had a mounting configuration that produced consistent image quality at speed.
Inference runs at 28ms per frame. The line runs at 120 units per minute. The maths works — but only with careful model quantisation and a dedicated inference pipeline that bypasses Python's GIL entirely.
