Industrial AI · Quality control

“Our customer finds the defect. Not us.”

Note on the shift board

Machine vision quality control

A camera above the line and a model trained on your own defects. It looks at every part the same way, in the first hour of the shift and in the eighth.

Send us 20 photos of a defect Book a Plant X-Ray

Inspection camera with a square LED light, pointing straight down over a roller conveyor carrying white PVC profiles.

unsure 0.64 · to a human

AI use in firms with 10+ staff, 2025: in rubber, plastics and non-metallic minerals, Greece 12.67% and EU 16.77%; in food, drink and tobacco, Greece 3.84% and EU 12.47%. Source: Eurostat.

Enterprises with 10+ staff, 2025. source (opens in a new tab)

The first category counts cement, glass and ceramics alongside plastics (NACE C22–C23).

Parts inspected: 12
Rejected: 4
Sent to a human: 1

The invisible problem

Your inspector is good. And human.

At 07:00 they look at every part. By 14:00 they look at whatever they can, because the line has sped up and the palletising is theirs too. The defects that slip through show up nowhere in production. They show up weeks later: in a credit note, a returned pallet, a buyer's email that opens with “Unfortunately”.

The cost scatters. A little in scrap, a little in re-inspection hours, a little in the discount the salesperson gives to keep the customer. No spreadsheet adds it up under the name “quality”.

If you sell to retail chains or large manufacturers, an escaped defect doesn't just cost a credit note. It costs your place on their supplier list.

How it works on your floor

From camera to reject gate.

1

What you already have

Often there is already an IP camera over the line. If not, an industrial camera and proper lighting: the lighting matters as much as the camera. We take the image over RTSP or GigE Vision, and a signal from the PLC for the moment each part passes.

IP camera · RTSP · GigE Vision · PLC trigger
2

The model

We train it on photos from your own line: good parts, and the defects that cost you. It runs locally, on a computer beside the line, so the decision lands within the cycle time and the image never has to leave the plant.

trained on your samples · runs on site
3 The decision, inside your system

Three paths, one record

“OK” and the part carries on. “Reject” and the PLC fires the reject mechanism. “Not sure” and the part goes to a human review queue, with its photo alongside. Every decision is logged with image, time and lot, and the quality lead sees it in the ERP or on the shift screen.

OK · reject · human review · ERP

Three-step diagram: a camera and light over the conveyor, a computer beside the line, and a fork where good parts continue, defective ones are rejected and uncertain ones go to a person.

Hardware: one camera and one light per inspection point, one computer beside the line. Often the camera is already there. We take care of the data, the models, the dashboards and the alerts.

What we measure

We measure what hurts your customer.

The formulas are the same for everyone. The values are yours, measured before the pilot.

Escaped defects (ppm)
The formula defects found by the customer ÷ parts delivered × 1,000,000
Scrap (%)
The formula parts rejected ÷ parts produced × 100
Inspection hours
The formula human inspection hours + re-inspection hours, per shift
Quality credit notes (€)
The formula Σ credit notes and returns marked “quality”, per month

The model's own marks

Missed defects
The formula defective parts passed as OK ÷ all defective parts
False rejects
The formula good parts rejected ÷ all good parts

We publish these two in the pilot report, not a bare “accuracy”. A model that rejects everything misses no defects. It also ships nothing.

How the benefit is worked out

Monthly benefit = (ppm before − ppm after) × parts per month ÷ 1,000,000 × cost per escape + inspection hours returned × €/⁠hour − running cost

Payback (months) = pilot cost ÷ (monthly benefit − monthly running cost)

€/hour: default €18.20 (Eurostat 2025, manufacturing). We replace it with your own. source (opens in a new tab)

The hours returned don't become redundancies. The inspector moves from “look at everything” to “judge the hard cases and find the cause”.

The pilot

6–⁠8 weeks. One line.

If the target isn't met, the report says so, with the same numbers we would have shown if it had been.

Before any of this, send us 20 photos from your phone. We tell you, free of charge, whether the defect shows up on camera and what lighting it would take. Send us 20 photos of a defect

NTX-PL-VQ6–⁠8 weeks

Pilot charter

Scope: One inspection point on one line or one mould family. The 2–⁠4 defect types that cost the most. An “OK, reject or human review” decision.
Duration: 6–⁠8 weeks
Phases: Weeks 1–⁠2Lighting, camera position, samples collected and labelled with your quality lead.
Weeks 3–⁠5Training, then a shadow run beside your inspector: the model decides, nobody acts on it yet.
Weeks 6–⁠8Live running with a human review queue. Escaped ppm, scrap, missed defects and false rejects measured. Decision.
Target: Example target: during the shadow run, the model catches at least as many defects as your experienced inspector, with false rejects below the limit your quality lead sets.
Cost: It starts with a Plant X-Ray, from €4,900, which includes a camera feasibility test on one line and is credited in full against the pilot.

Signed before any code is written.

Evidence

Play first. Read second.

Try it yourself
60 seconds, an AI-made image of a bottling line, eight defects. Then see what the machine found.
Play “Spot the defect”
1 defect type · 340 images
That was the whole scope of a 2024 lab study on injection-moulded parts: one defect type (burrs at a hole), 340 test images, one plant. It reports accuracy close to 100%. We don't show that as proof: real lines have glare, dust and hurry. That's why we measure on yours.
source (opens in a new tab)
12.67% · 16.77%
of firms with 10+ staff in rubber, plastics and non-metallic minerals (cement, glass, ceramics) use AI, in Greece and the EU (2025).
source (opens in a new tab)
3.84% · 12.47%
of firms with 10+ staff in food, drink and tobacco use AI, in Greece and the EU (2025). In 2024 Greece stood at 9.75%, above the EU: in small samples the share swings a lot from year to year.
source (opens in a new tab)

Our own measured project

We haven't published a measured vision project yet. The first one could be on your line.

Honest limits

Where it gets it wrong. And what we do about it.

Every vision model makes two kinds of mistake: it rejects good parts and lets bad ones through. We don't hide them. We tune them, and we measure them in front of you.

Show one of the model's mistakes and how we handle it

One of its mistakes, in full

What the model said: The model saw a crack, at 0.62 confidence.
What actually happened: It was a reflection of the ceiling light.
What happened next: Below the 0.80 threshold, the part went to human review. The inspector passed it as OK, the photo joined the retraining set and a diffuser went on the light. The mistake was counted in that week's false rejects.

A confidence threshold: below it, a person decides.
A human review queue for the grey zone, with the photo alongside.
A record for every decision: image, time, lot, confidence.
Retraining whenever the material, mould, lighting or supplier changes.

What it needs from you

Samples: a few dozen to a few hundred photos per defect type, depending on how visible it is. We collect them together during the pilot.
A fixed part position, or a known conveyor speed.
A quality lead who decides what counts as a defect. The model learns their criteria, not ours.

What it won't do

It can't see what a camera can't: internal stresses, dimensions below its resolution, smell, taste.
It doesn't replace the lab tests or dimensional checks your customer requires.
It doesn't promise 100%. No vision system achieves it, and whoever promises it has never stood beside a line at 14:00.

Play

One inspector, one camera, one shift.

The belt runs at line speed. Choose who's watching.

Fully playable by keyboard.

Skip the demo: read how it works

Line speed

Parts
Defects caught
Defects missed
Sent to a human

Out of 150 parts, the inspector caught 1 of 6 defects. The camera caught 4, missed 1 and sent 4 to a person.

The line doesn't slow down at hour seven. Neither should the inspection.

Use case: here the “inspector” looks at one part in fifty, a rate set for this example. The camera makes mistakes too, as every camera does.

The demo didn't load. The text below says what you would have seen.

In the demo, a conveyor carries plastic parts at line speed. With the “inspector at hour seven”, one part in fifty is checked, so most defects pass unnoticed. With “AI on every part”, every part gets a box and a decision; most defects are caught, one slips through, and the uncertain ones go to a person.

What you just saw

Mechanism: The same attention on every part; a person only in the grey zone.
Why it works: Fatigue isn't negligence. It's biology. The camera doesn't tire, but light can fool it. That's what the review queue is for.
Data it needs: Photos of good and defective parts from your line, a PLC trigger, your quality lead's criteria.
Honest range: How much a model catches on your line we measure in the shadow run, beside your inspector. Not before.

Send us 20 photos and we'll tell you.

Send us 20 photos of a defect

From your phone, as they are. Ideally ten showing the defect and ten good parts. We tell you whether the defect shows up on camera and what lighting it would need. No charge, no commitment.

We don't use them to train anyone else's models. We don't show them to anyone.

Same logic, another trade

Every trade has its own defect.

Same mechanism: a camera, a model trained on your samples, a person in the grey zone. Six lines where it works.

Use cases: the same mechanism at work in other trades.

Honey packer, Crete