The defect you find after dispatch costs four times the one you find on the line

Every plant manager knows the cost curve, even if nobody has written it down. A defect caught at the station where it was made costs a few minutes. Caught at final inspection, it costs the work done on it since. Caught by the customer, it costs the unit, the shipping, the replacement, the credit note, and a conversation you did not want to have.

The problem is that catching it at the station requires somebody to look at every unit. Nobody looks at every unit. They look at a sample, and they look less carefully at four in the afternoon than at nine in the morning, because they are human beings and it is the fourth hour of looking at the same thing.

What a camera changes

Computer vision, in a factory, is not a robot. It is a camera pointed at a fixed spot on the line, and software that has been shown enough examples of "correct" and "wrong" to tell them apart on units it has never seen.

What it changes is not accuracy on any single unit — a good inspector, fresh, beats it. What it changes is coverage and consistency. It checks unit 4,000 exactly as carefully as unit 1, at 4pm exactly as at 9am, on a Sunday night shift exactly as on a Monday morning. Most quality problems are not caused by an inspector being bad at their job. They are caused by there being 4,000 units and one pair of eyes.

The typical applications, in rough order of how often we are asked for them:

• Surface and finish defects. Scratches, cracks, contamination, colour deviation, incomplete coating.
• Presence and completeness. Is the cap on, is the label there, is it straight, is the right number of items in the box before it is sealed.
• Counting. What actually came off the line, versus what the shift report says came off the line.
• Reading. Batch codes, expiry dates, labels — checking that what is printed matches what should have been printed.

That last one quietly saves more money than the first, in food and pharmaceutical plants especially. A whole batch with the wrong expiry printed is not a quality problem, it is a recall.

Where it fails, and why most projects die there

We would rather tell you this before you buy than after.

Lighting kills more vision projects than algorithms do. A model trained on morning light in a shed with an open door will quietly degrade at four in the afternoon and fall apart in the rainy season. Fixed, controlled lighting at the inspection point is not an optional extra; it is usually the difference between a system that works and one that gets switched off in month three.

Camera placement is a one-way door. If the defect is only visible at an angle the camera cannot see, no amount of software fixes it. This is worth an hour of argument at the start.

The examples have to come from your line. Not a public dataset, not the vendor's demo footage. Your product, your lighting, your packaging, your failure modes. Collecting a few hundred examples of the defects you actually get is real work, and it is work that has to happen while production continues.

And accuracy is two numbers, not one. A system that catches 99% of defects while also flagging 8% of good units is not a good system — it is a system your operators will start ignoring within a fortnight, and an ignored alarm is worse than no alarm. Ask any vendor for both numbers. If they only give you one, that tells you something.

The bandwidth question, which matters more here

Most vision products assume a data centre. They stream video out, process it somewhere else, and send an answer back. On an industrial estate outside Douala, with the connection most plants actually have, that is a fantasy — and the moment the link drops, your quality system stops.

So we run this on site. A small box next to the line, doing the detection locally, in milliseconds, with no dependency on the internet being up. What leaves the site is the result — counts, defects, alerts — not the video. That is cheaper, faster, more private, and it keeps working during an outage.

How to scope a first project that finishes

The vision projects that fail are the ambitious ones. The ones that succeed pick a single defect type, on a single line, at a single station, and prove the number.

A sensible first project looks like this:

1. Pick the defect that costs you the most, not the one that is most interesting.
2. Fix the lighting and the camera position before writing a line of code.
3. Collect examples for two to four weeks while production runs normally.
4. Run the system in parallel with your existing inspection, without acting on it, until you can compare the two honestly.
5. Only then connect it to anything that stops a line or rejects a unit.

Step four is the one everybody wants to skip, and it is the one that determines whether your team trusts the system afterwards.

What we would ask you first

• Which defect costs you the most, counting rework, credit notes and lost customers?
• Where in the line is it first visible to a camera?
• What is the light like at that point, at 4pm, in August?
• What happens today when a defect is found — who is told, and how fast?

If you can answer those four, a first vision project is a matter of weeks rather than a research programme.