Blog
The factory that inspects by eye, and what AI can actually see instead
Visual inspection is the fastest place for AI to earn its keep on a production line. It is also the easiest place to start badly.
Published
Two twenty-five in the morning, night shift. An inspector watches parts go past on the belt, one at a time. She has done this for six years, and her eye catches a scratch under a millimetre better than anyone who joined three months ago. By four in the morning, though, her eye is the eye of someone who has been awake for eighteen hours.
Every plant manager knows this problem. It is not that the inspectors are not good. It is that the quality of inspection is not constant — across a shift, across a week, across a career. It moves with fatigue, with the light on the floor, and with how fast the line is running on a week when orders are heavy.
Why inspection is the natural first project
Of all the places AI can help in a factory, visual inspection fits best, because it has three properties at once.
It is a repeated judgement against a fixed standard. Pass or fail; this mark is acceptable, that one is not. That is exactly the shape of problem an image classifier is good at.
Its results are measurable immediately. Parts that reached a customer and came back are a number the quality department already keeps. Nobody has to invent a new metric and then argue about it.
And it does not require tearing anything else out. A camera and a light go in at the inspection station without touching the ERP, without touching the MES, and without stopping the line to change a process.
What it does, and what it does not
A visual inspection system that actually works is a fixed camera, controlled lighting, and a model trained on images of real defects from that line — not on sample images from the internet.
It is good at defects that recur: scratches, chips, misalignment, missing components, colour drift. It gets better as more defect images accumulate.
It is not good at defects it has never met. If a supplier changes material and a new kind of flaw appears, the model will pass it. That is not a shortcoming of the technology; it is what learning from examples means. It is also why a system like this needs a person in the loop rather than an installation and a handshake.
Where these projects fail
No defect images. A plant with good quality control has few defects, which means few images to learn from. The answer is not to wait for defects to accumulate. It is to start capturing images systematically on day one, before there is any model at all.
Unstable light. A great many projects die here rather than at the algorithm. An image taken in the morning and one taken at night, under different amounts of daylight through a window, are two different datasets as far as the model is concerned. An enclosure and controlled LED lighting pay back more than a better model does.
Tuning to the wrong side. A system tuned to catch every defect will reject so much good product that the operators stop believing it within a fortnight. A system tuned to bother nobody will let defects through. Choosing the balance is a business decision, not a configuration setting, and quality should be the one making it.
How to begin without wasting a year
Pick one part, one station, and the two or three defect types that occur most. Capture images for a month or two alongside normal human inspection, then compare the two quietly before the system decides anything on anyone's behalf.
That comparison period matters more than most people expect. It is when the quality team sees how the system fails, and decides for itself how far to trust it. The trust built there is what keeps the system in use in its second year.
What GIPSIC does here
Work like this does not end at the model. It needs cameras and lights mounted correctly, hardware that survives a factory floor, a system that stores images and decisions so they can be revisited, and a screen a shift leader can actually open without calling IT.
Our hardware and software people sit in the same team — device design, embedded firmware, the cloud platform behind it, and certified installers on site — so camera placement and database structure get discussed in the same conversation.
And if we conclude your problem should not start with AI at all, we will say so and explain why. Some plants get more out of making their inspection records digital first, and thinking about models the year after.
If you would like to talk through your line, get in touch.
About the author
Film — Wisit. A businessman who still does his own BA work more often than he probably should, and writes a fair bit of code, front and back. Runs two or three small businesses. Follows technology and business obsessively, in Thailand and everywhere else. Off the clock: physics, astronomy, DIY, and anything to do with networks. Music always on, though he cannot sing. Plays instruments anyway, badly. Plays a lot of sport, racket sports above all. Not much of a traveller by himself, but happy to take Mint anywhere in the world. A man who fears — sorry, loves — his wife. One flaw: he barely touches video games.
Written with Claude Opus 5