What the line lead knows now stays on record

The problem

The person who first notices a drift in formation has no free hand to write it down.

Critical operational knowledge lives in the line lead's head and never enters any system, even though he's the first to notice drift in a sensitive chemical process like formation or curing.

• The assembly line lead spends the shift walking between grid casting, formation and the cast-on-strap welding station, with his hands full most of the time. Formation, curing and welding each have their own rhythm, and he is the one stitching them together.
• He is usually the first to notice drift in a sensitive chemical step: acid temperature climbing in formation, a curing load running long, a COS welder leaving cold joints. Experience tells him something is off long before any report shows it.
• None of that enters any system, because he cannot type while adjusting a welder or holding a tray of plates. Walking to a terminal means leaving the line, so the observation stays verbal or stays in his head.
• At shift end that knowledge dies unrecorded, and the next lead runs into the same drift without knowing it was already seen. Over months, the same issues are rediscovered shift after shift and never reach a root-cause analysis.

How it fits the IRIS system

Connect in voice mode with a custom wake word — he dictates while walking the line.

Connect voice mode with a Bluetooth headset and a custom wake-word ("iLEAN, note this"). The LLM structures the dictation as an incident linked to the active lot/line/shift, and returns a reply by voice or push notification.

Only what follows the wake word is recorded; nobody else on the floor is listened to. In return the line lead gets support back in seconds, by voice or push, instead of one more report at the end of a long shift. Shift handover stops being a hurried conversation at the door and becomes a list the incoming lead can read.

See the full IRIS architecture →

Before and after

The battery line lead's shift today versus the shift with a headset

Aspect	Today	With iLEAN Connect
Acid temperature drift seen in formation	Mentioned at handover, if remembered	Incident on the active formation lot
Cold joint noticed at the COS station	Told verbally to maintenance	Structured incident with welder and lot
Hands needed to log it	Both, and a walk to the office PC	None: wake word and voice
Link to formation lot and welding station	Rebuilt later from memory	Attached the moment he speaks
Help from maintenance or quality	Phone calls and walking around	Reply by voice or push in seconds
Knowledge at shift handover	Goes home with him	Available to the incoming lead

Knowledge that dies at shift end → enters the system in real time and gets cross-referenced with live formation/assembly data.

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.

• No standalone payback: this is an enabling piece, and we do not put a month range on it because its value shows up in the other cases.
• It is hard to monetize upfront, but it underpins every continuous-improvement plan in the plant: without recorded incidents there is nothing to analyze.
• Knowledge that used to die at shift end enters the system in real time and is cross-referenced with live formation and assembly data.
• We estimate its effect together with your continuous-improvement lead, on the incidents that start appearing in the record.

Hard to monetize upfront, but it enables every continuous- improvement plan in the plant. *Figure to validate* with the continuous-improvement lead.

And the fair question from the production manager

“Is this a way to listen to the operators?” — no: it only records what follows the line lead's wake word. Turning a spoken note into a structured incident with lot, line and shift is an anchored task, where the best models stay below 1.5% error [1]; and the line lead hears the summary back and can correct it before it is stored.

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