AI quality inspection on your own criteria

See rejects before they leave the building

AI quality inspection, in industry also called quality control, is the automatic assessment of every product in the production line with an image recognition model. Unlike a sample check, the system assesses everything, consistently and at the pace of the line.

At Twentynext that means: a model that assesses images from your production line, trained on your own products and your own rejection criteria. Not a generic model that recognises ‘defects’, but one that knows what counts as a reject for you.

What the system does

It assesses every product while the line is running:

Consistently: the same deviation gets the same judgement at the end of a night shift as at the start.
Including deviations that are hard to catch with the naked eye: hairline cracks, colour shifts, misalignment.
Without affecting throughput; the assessment happens at the pace of production.

From reject to cause

Every judgement is also data. Once there is enough of it, the question shifts from ‘which product do we reject’ to ‘why does it go wrong’:

Which combination of settings, batch and time precedes a spike in rejects.
Which small deviations signal that something in the line is drifting, before the product falls outside the standard.
Where in the process the cause occurs, rather than where it becomes visible.

What it delivers

The value rarely lies in the inspection itself, but in what disappears around it: fewer products rejected late in the process, fewer manual sample checks, and less discussion about where the standard lies.

What it needs

An honest answer up front prevents an expensive surprise halfway through. A working system requires:

Images of sufficient quality. Stable lighting and a fixed camera setup matter more than a more expensive camera.
Examples of rejects. The model learns from your own rejected products. If there are hardly any, the project starts with collecting them.
A place in the process. The judgement has to land somewhere: a signal to the operator, an ejection mechanism, or a record in your existing systems.

The connection with existing production equipment, IoT platforms and reporting is part of the work, not a separate phase afterwards.

What you get

Not a package, but a working solution

Twentynext is not a software vendor. We examine the problem, design the solution and build it until it is running in production.

What we deliver

A solution that runs in your own process, on your own data. No licence for someone else’s software, and no report with recommendations.

Where we start

With the business question, not with the model. Which decision needs improving, and what information does it depend on? Technology is the last step, not the first.

Who builds it

The same people who design it. From data analysis and architecture to production and maintenance, without handing over to another party.

How we approach a project

Solutions we have built

What else can we do for your organisation?

A few solutions we have built before:

Data and reporting platform

Microsoft Copilot implementation

Population dynamics

Product recognition

Quality inspection

Frequently asked questions

What does AI quality inspection do in a production line?

A model assesses images from your line against your own rejection criteria, after being trained on your own products. It makes judgements at production speed, without affecting throughput, and gives the same judgement on the same deviation at the end of a night shift as at the start.

What is needed to get started?

Three things: images of sufficient quality, with stable lighting and a fixed camera setup mattering more than an expensive camera; examples of rejects from your own production; and a point in the process where the judgement is used: a signal to the operator, an ejection mechanism or a record in your existing systems.

What if we have kept hardly any rejected products?

Then the project starts by collecting them. That does not rule out a project, but it does determine the lead time: the model learns from your own rejects, not from a generic model that recognises defects.

Does the system also say why things go wrong?

Every judgement is also data. Once there is enough of it, the question shifts from which products you reject to why things go wrong: which combination of settings, batch and time precedes a spike in rejects, and where in the process the cause lies rather than where it becomes visible.

Does Twentynext supply a ready-made inspection package?

No. We are not a software vendor. You get a solution that runs in your own process on your own data: no licence for someone else’s software, and no report with recommendations.

Does it connect to our existing equipment?

The link with existing production equipment, IoT platforms and reporting is part of the work, not a separate phase afterwards. The judgement has to be captured somewhere to have value.

Is quality control the same as quality inspection?

In practice, yes: both words describe assessing products against a standard, and industry and vendors use them interchangeably. On this page we mean the same thing by both: automatic inspection of every product in the line, against your own rejection criteria, instead of a manual sample check.

Does AI quality inspection also work in the food industry?

Especially there. Natural products vary in colour, shape and size, and precisely that variation makes fixed inspection rules unreliable. A model trained on your own product stream learns what normal variation is and what a deviation is. The same goes for every sector with high volumes and visual standards.

What does AI quality inspection cost, and when does it pay for itself?

An initial set-up with a camera, lighting and a model usually costs between €25,000 and €75,000, depending on the number of product variants. Payback comes from rejects the customer no longer sees and from inspectors spending their time on exceptions rather than inspecting everything on the belt. For a line with more than 1 percent rejects, that is usually within a year. These are guidelines from our projects, not a guarantee: every situation calls for tailored work, and we determine costs and lead times for each organisation.

Why is implementing a vision model harder than it looks?

The model is the smallest part of the challenge. Lighting that changes with the seasons, products positioned slightly differently and a line that cannot stop all determine whether it works. That is why we start by collecting images on the real line for a week, before any model is trained.

How does this compare to manual inspection?

Manual inspection remains necessary for the cases the model is unsure about. The model removes the volume, while the person assesses the uncertain cases. The percentage of uncertain cases is the metric we steer by.