Clean recycled aggregate, guaranteed by computer vision

The problem

At belt speed, the human eye is sampling, not inspecting.

At belt speed, the human eye can't systematically catch wood, plastic or scrap-metal fragments in the recycled aggregate. If that contaminated material reaches a customer or the site reusing it, it compromises product quality and its CE marking.

• Demolition feedstock arrives mixed by definition: concrete and masonry, but also timber, plastic sheeting, gypsum board and scrap metal that makes it through the crusher intact.
• On the output belt, somebody glances at the flow now and then between other jobs. That is sampling with gaps, and the fragment that slips through is precisely the one nobody was looking at.
• When contaminated aggregate reaches a customer or a site reusing it, the complaint is never about one bucket: it questions the whole batch, its grading report and the CE marking behind it.
• And the cost is rarely the material itself. It is the relationship with the customer who chose to take your recycled aggregate instead of buying virgin, and who now has a reason not to.

How it fits the IRIS system

Edge — local CNN vision over the output belt, with no cloud round-trip.

Edge — local CNN vision, no cloud round-trip, trained on clean and contaminated examples of the specific material the plant processes. In milliseconds per belt section, it triggers diversion of non-conforming material.

The model is trained on clean and contaminated examples of the material this plant actually processes, not on a generic dataset bought off a shelf. Demolition feedstock changes with every building taken down — an office block is not a warehouse and neither is an old industrial plant — and that variation is exactly what the training has to cover.

See the full IRIS architecture →

Before and after

Today's visual check versus Edge vision on the output belt

Aspect	Visual inspection	With iLEAN Edge
Coverage	Occasional glances	The whole belt, continuously
Timber and plastic sheeting	Slip through at belt speed	Classified belt section by section
Gypsum fragments	Hard to spot by eye	Trained explicitly
Point of detection	The customer's complaint, weeks later	Before it ever joins the stockpile
Non-conforming material	Already inside the certified batch	Diverted in milliseconds
Borderline cases	Judged by whoever is there	Escalated to a person

Before: batch sampling with gaps only caught when a customer complains. After: continuous inspection of the entire belt.

Impact estimate

Impact estimate — to be validated with your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 5-12 months, depending on the volume of material the plant processes over a year.
• Protection of the recycled aggregate's CE marking, which is what allows the material to be sold or reused on another site at all.
• From batch sampling with gaps to continuous inspection of the entire belt, with no change to the crushing plant, its screening decks or the crew running them.
• And protection of the relationship with customers who reuse your aggregate, which is the part a single rejected load actually damages.

Estimated payback of 5-12 months depending on volume processed, protecting CE marking and relationships with customers reusing the aggregate. Estimate to validate.

And the fair question from the production manager

"What if it diverts good material?" — the false positive is the real risk of any vision system, and on an output belt it is paid for in tonnage. That is why the model is trained on your own feedstock and under that belt's actual lighting, and why a borderline fragment is escalated to a person instead of being dumped into the reject bin by default. The classification itself is an anchored task, where the best models drop below 1.5% error [1], and every human decision on a borderline case feeds back into the training set.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.