Contaminants pulled off the sorting belt

The problem

The picking cabin cannot see everything the belt carries.

Fully manual sorting is impossible to do without gaps at the belt's real pace. Partial visual sampling leaves holes that translate into lower-quality aggregate, and contaminants (plastic, wood, plaster) can make an entire batch fail the required grading.

• In the picking cabin, sorters pull wood, plastic, gypsum and metal off the belt by hand. At the belt's real pace, that is sampling, not sorting.
• What slips through goes straight to the crusher and ends up inside the recycled aggregate, where it can no longer be separated. The crusher turns a removable piece of wood into fines that stay in the product.
• Gypsum is the most treacherous: a few kilos of gypsum or wood are enough for a whole batch to fail the required grading and contaminant limits, and be downgraded or sent to landfill, with the gate fee paid on top of the lost sale.
• And the sorters' attention drops over the shift, exactly when the heaviest mixed loads tend to arrive. Nobody blames the sorters; the pace is simply beyond the human eye.

How it fits the IRIS system

Edge — local vision over the sorting belt, inferring per segment in milliseconds.

Edge: local CNN vision, with no need to send images to the cloud (latency, cost, bandwidth), trained on the plant's own waste flow. It infers per belt segment and flags or rejects the contaminant before it reaches the crusher.

Inference runs at the plant, in a box next to the belt, not in the cloud: no latency, no bandwidth bill, no dependence on the yard's network. The model is trained on your own waste flow, so it knows your concrete, your brick and your contaminants. Every reject is logged against its batch, so the contaminant count is no longer an estimate.

See the full IRIS architecture →

Before and after

Picking by eye versus Edge vision ahead of the crusher

Aspect	Today	With iLEAN Edge
Share of the flow inspected	Partial, depends on the sorters	100% of the belt
Wood and plastic film	Missed when the belt is loaded	Flagged per segment
Gypsum and plaster fragments	Hard to tell from mortar	Classified by texture and color
Where it is caught	After the crusher, in the batch	Before the crusher
Batches rejected for impurities	A recurring loss	Drastically fewer
Images sent to the cloud at a C&D recycling plant	—	None: inference on site

Partial visual sampling → continuous control over 100% of the belt's flow. Aggregate batches rejected for impurities → drastic reduction.

Impact estimate

Impact estimate — to be validated with your volume and aggregate price.

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 processed volume and the sale price of the aggregate, which sets what every degraded batch costs.
• Fewer recycled aggregate batches degraded or rejected for impurities, which is where most of the value sits.
• Continuous control over 100% of the belt's flow instead of partial visual sampling, with a contaminant count logged per batch.
• And the sorters' attention goes to the borderline cases the camera escalates, not to everything that passes.

Estimated payback 5-12 months depending on processed volume and aggregate sale price, fewer degraded batches. *Estimate to validate*.

And the fair question from the production manager

“What if it rejects good concrete?” — the false positive is the real risk, which is why the model is trained on your own waste flow and lighting, not on a generic catalog. Telling plastic film or wood from mineral material is a well-bounded visual task — the same kind of anchored problem where the best models drop below 1.5% error [1] — and uncertain objects, the ones the model is not sure about, are flagged to the cabin instead of being ejected. Every decision, right or wrong, feeds the next training round.

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