AI quality control for Queensland manufacturers: the camera is the easy part

Every vendor selling manufacturing AI shows you the same thirty seconds. A camera watches parts go by, a red box appears around a cracked one, everyone in the boardroom nods. It looks like the problem is solved.

The camera is the cheap part. It’s also the part that fools people into buying a system that never pays for itself. Because a red box on a screen isn’t quality control. It’s a red box. Quality control is everything the demo skipped: what counts as a defect, what the system does when it’s wrong, and what actually happens when the box appears.

Computer vision can cut defects, rework and warranty claims, and it’s one of the better first applied AI projects a manufacturer can take on, precisely because you can measure whether it worked. But only if you build the boring 90 per cent the demo left out.

“Improve quality” is not a spec. This is.

The fastest way to waste money here is to start with a goal too vague to build against. “Improve quality” tells the system nothing. Scratches, cracks, wrong labels, missing fasteners, colour drift, a part seated wrong, a packaging error: those are completely different problems, they fail in different ways, and lumping them together gets you a system that’s mediocre at all of them.

Take a fabrication shop on the Darling Downs that supplies brackets to a mining equipment OEM. Ask the QA lead what hurts and she won’t say “quality”. She’ll say the same joint on the same assembly keeps getting through with incomplete welds, the customer’s receiving inspection catches maybe half of them, and every one they catch triggers a credit, a corrective action report and a tense phone call. That’s not an aspiration. That’s one defect, on one part, at one station, with a cost attached, and it’s exactly the shape a first vision project should have.

“Flag any assembly leaving this station with a missing fastener, before it’s packed” is something you can build, measure and trust. So is “catch surface marks above two millimetres on this face of this component.” Pin it down that tightly. You can add the second defect once the first one is paying for itself. Trying to catch everything at once is how you catch nothing reliably.

The false reject is the cost nobody quotes

Here’s the number the demo never mentions, and it’s the one that decides whether staff keep the system switched on. Every detection system makes two kinds of mistake, and they cost you in opposite directions.

A missed defect (the bad part it waves through) turns into rework, a warranty claim, or an angry customer three steps downstream, where it’s far dearer to fix than at the station. A false reject (the good part it bins) is money straight in the skip, plus a line that stops for nothing, plus operators who learn to distrust the machine. Tune the system to catch every possible defect and you’ll drown in false rejects and your staff will override it by lunchtime. Tune it to never cry wolf and it’ll miss the ones that matter.

Put rough numbers on it and the trade-off stops being abstract. Say a line runs 4,000 units a shift and the system false-rejects 2 per cent. That’s 80 good units a shift heading to the reject bin or looping back for a second look, every shift, indefinitely. If a unit is worth $30, the machine is quietly costing you around $2,400 a shift in good product unless someone re-screens the rejects, in which case you’ve hired the inspection job straight back. Now run the same sum on the other side: if one escaped weld costs $600 in rework, freight and credits, the arithmetic tells you which way to lean for this particular part. On a cheap high-volume part you tolerate a few more misses. On a safety-critical one you wear the false rejects and staff the review queue properly.

That trade-off isn’t a bug to engineer away. It’s a business decision about which mistake hurts you more on this particular part, and it belongs to you, not the vendor. A system that can’t show you its false-reject rate on your own parts, under your own lighting, is a system you can’t cost.

Your factory is not a photography studio

Vision models trained on clean, evenly lit sample photos fall apart on a real floor, and this is where a lot of these projects quietly die. Your factory has glare off metal, dust on the lens, parts moving past at speed, shadows that shift as the sun tracks across the roller door, and a different operator loading the jig on each shift. The model has to have seen all of that, or it’ll flag the lighting as a defect and pass the actual crack.

So the image set has to come from the floor, in the conditions the system will actually work in, including the borderline cases that are hard to call either way. Those edge cases are exactly where the review queue does its real work, so the model needs to have met them in training, not for the first time in production. Lab-perfect photos are worse than useless here, because they give you a system that demos beautifully and fails on Tuesday.

And the floor keeps changing after launch. A new supplier batch arrives with a slightly different surface finish and the reject rate triples overnight. Winter light through the western windows shifts every image the camera captures between May and August. The line speed goes up ten per cent and motion blur creeps in. None of these are exotic failures. They’re the ordinary drift of a working factory, and the system needs a plan for them: someone watching the reject rate for sudden jumps, a routine for feeding fresh examples back into training, and an agreed trigger for when the model gets a refresh instead of a shrug.

Detection has to trigger an action, or it’s theatre

A detection event that ends with a red box on a screen has moved the work, not removed it. Someone still has to see the box, decide what it means, and do something. That “someone still has to” is most of the job you were trying to automate.

The value is in the wiring behind the detection. The box appears and the line stops, or the item diverts to a reject bin, or the photos attach themselves to that batch record, or a supervisor gets a message, or the defect count lands in the production report on its own. That integration layer, connecting the camera to the systems that actually run the factory, is where the project gets real and where most of the engineering actually is. It’s also the least glamorous part, which is why the demo skipped it and why it’s the part worth paying for.

Route the uncertain calls to a person early on, and let the business learn which detections are reliable enough to act on automatically and which still need eyes. Every case a person rules on is a labelled example that makes the next month’s model better. The review queue isn’t a stopgap until the AI is good enough. It’s the thing that makes the AI get good.

Ask who owns the images before anyone trains anything

Six months in, you’ll have thousands of labelled photos of your parts, your defects and your borderline calls, each one judged by your best QA people. That library is the most valuable thing the project produces. It’s also the thing some vendors quietly keep.

Read the contract for where the training data and the trained model live. If the images sit in the vendor’s cloud, in their format, and the model can’t leave their platform, then every future improvement, every new SKU and every price rise happens on their terms. That’s vendor lock-in with a manufacturing accent, and it’s far cheaper to negotiate before the first camera is bolted above the line. You want the images, the labels and ideally the model itself exportable in a standard format, and you want a straight answer on what retraining costs when a product changes, because products always change. If the parts themselves are commercially sensitive, running the whole loop on infrastructure you control is worth pricing too.

Start where it already hurts

Don’t wire up the whole factory. Pick the one defect that costs you every single week, the one you could name right now, and prove the loop end to end on that: capture the image, classify it, handle the false rejects sensibly, trigger the action, and put a number in the production record. Prove it pays for itself on one station before you touch a second.

A scoped first station like that is typically a five-figure project, not the seven-figure floor-wide platform the glossier vendors pitch. If the defect you’re chasing burns a few thousand dollars a week in rework and credits, the payback window is months, and you can check it against the production record within a quarter. If nobody in the business can name a defect with a weekly cost attached, that’s an answer too: vision probably isn’t your first AI project.

The test is never whether the demo looked clever in the boardroom. It’s whether fewer bad units make it down the line next quarter, and whether the production record becomes something you’d trust in front of a customer. If you’ve got a defect that bleeds money every week, tell us the part and how it fails and we’ll scope a first station. Or start with the AI readiness assessment to see whether vision is even the right first move.