Critical site-supervisor alerts, in seconds

The problem

The channel that carries the most urgent news is the one nobody is watching.

Critical alerts from third parties — the site supervisor, the authorities, the developer — arrive by email or WhatsApp to someone who can't check their inbox in real time from the field. By the time it's finally read, the machine may have been working the affected zone for hours, with the environmental or safety risk that implies.

• A finding appears — an unmapped service, a structural element that was not in the survey, suspected asbestos-containing material — and the site supervisor writes an email ordering a stop in that zone.
• The developer, the authorities and the engineer of record all use that same channel, and the message lands in an inbox belonging to somebody who is out on a site, in a helmet, with no realistic way to read it in the next couple of hours.
• One to three hours of human latency is normal, and during those hours the machine keeps working the zone the order was meant to freeze.
• When it is finally read, the question is no longer what to do but what has already been demolished, and the answer has to be given to the supervisor in writing.

How it fits the IRIS system

Connect in chaotic-sources mode — a dedicated mailbox, silent copy, around the clock.

Connect's chaotic-sources mode: a dedicated mailbox in silent copy 24/7. A grounded LLM extracts the alert's intent (finding, stop order, scope change) and entities (location, zone, deadline), automatically cross-referencing them with the active demolition plan and the machines working that zone, returning the concrete action to the affected foreman.

Finding out sooner is only half of it. What changes the outcome is the cross-check: the alert is matched against the active demolition plan and the machines working that zone, so what comes back is not a forwarded email but a concrete action, addressed to the person holding the radio. On a site, an instruction that names the machine and the zone is acted on; an email that has to be interpreted is not.

See the full IRIS architecture →

Before and after

Today's alert versus the cross-checked alert

Aspect	Today, in the inbox	With iLEAN Connect
When the field finds out	1-3 hours later	Seconds
Who receives it	Whoever it was addressed to	The foreman of the affected zone
What arrives	An email to interpret	The action, the machine and the zone
Work in the frozen zone	Continues until someone reads it	Stops on the alert
Scope of the stop	Widened to be safe	Bounded to the zone named
Trace of the notice	Buried in a thread	Logged with its timestamp

Before: 1-3 hours of human latency until the foreman finds out. After: seconds, with a directed instruction naming the specific machine and zone.

Impact estimate

Impact estimate — this one is defended as exposure avoided, not as hours saved.

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.

• The brief puts no payback range on this case and neither do we: the risk avoided does not reduce to a single figure.
• What it does change is measurable: from 1-3 hours of human latency down to seconds, with a directed instruction naming the machine and the zone.
• The exposure it protects is environmental and safety-related — the two kinds that stop a site, reach the developer and end up in a report.
• It is a case to size with your safety officer, against the last stop order that was read too late.

Risk avoided is hard to pin to a single figure, but critical in terms of environmental or safety exposure avoided. Estimate to validate with the safety officer.

And the fair question from the production manager

"Are you going to read our company email?" — no. iLEAN sits on silent copy of a dedicated mailbox used with the site supervisor, the developer and the authorities: it sees nothing else. On extraction, classifying the intent of a notice — finding, stop order, scope change — and pulling out zone and deadline is an anchored task, where the best models drop below 1.5% error [1], and the action is always executed by a person.

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