The aggregate crushing plant, digitized without touching it

The problem

The plant that makes your product is the one that records nothing.

The crushing plant's panel constantly shows metric tons produced, moisture and consumption — vital data that dies on screen every time the batch changes, because no one transcribes it systematically. Without that correlation, figuring out why a batch of aggregate came out with more fines than expected means reconstructing what happened at the plant that day from memory.

• The plant producing recycled aggregate is often a decade or more old, with a proprietary panel, a closed protocol and no accessible network port. Replacing it is hundreds of thousands of euros and weeks with no aggregate coming out.
• Its display shows, continuously, metric tons produced, moisture and consumption — and every one of those readings dies on screen when the batch changes, because nobody transcribes them systematically.
• So when a batch comes out with more fines than expected, nobody can tell whether it was the feed material from that particular demolition, the closed side setting left from the previous job or a screen deck about to blind.
• The CE marking of the recycled aggregate asks for traceability of what the plant produced, and today that traceability is a stack of notes somebody wrote down on the days they had a spare minute.

How it fits the IRIS system

Connect in photo-on-panel mode — solved from outside the cabinet.

Connect's photo-on-panel mode: a fixed industrial camera aims at the display and an OCR pipeline with a grounded LLM interprets each reading, structuring it by batch and shift in central memory, without touching the original electrical cabinet.

Nothing is installed inside the plant, no port is opened and its manufacturer's warranty is untouched. On a machine the whole site's aggregate output depends on, that is the difference between a case that gets approved in a week and a project that waits forever for a shutdown window nobody wants to give it. The plant carries on crushing exactly as it is; the only thing that changes is that what it produces now leaves a trace.

See the full IRIS architecture →

Before and after

Today's plant versus the plant being read

Aspect	Today	With iLEAN Connect
Output per batch	On screen, then gone	Digitized batch by batch
Moisture and consumption	Never correlated	Tied to the batch and the shift
A batch with excess fines	Explained from memory	Explained from the readings
CE-marking traceability	Loose notes	History per batch
The electrical cabinet	—	Untouched, warranty intact
Plant replacement capex	In the plan	Indefinitely postponable

Before: zero output-quality correlation, data dying on screen. After: per-batch analysis in minutes, and plant replacement becomes indefinitely postponable.

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 5-10 months, counting the avoided plant-replacement capex first, which is the largest single line in the case.
• Per-batch analysis in minutes instead of a reconstruction, weeks later, of what somebody thinks happened at the plant that day.
• Aggregate traceability improved exactly where it is demanded: the CE marking of the material you sell or reuse on another site.
• And the replacement decision stops being forced by the fact that the plant reports nothing: the machine crushes perfectly well, it simply never told you what it was doing.

Estimated payback of 5-10 months, avoiding plant-replacement CAPEX and improving aggregate traceability for CE marking. Estimate to validate.

And the fair question from the production manager

"Can a camera pointed at a screen be trusted for a traceability record?" — here the task is anchored: a fixed display, fields in known positions and physically bounded ranges, where the best models drop below 1.5% error [1]. Any frame whose value falls outside the plant's possible range is discarded instead of accepted, which is exactly what would invalidate the series in front of a certification body.

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