Heat startup in the melt shop, digitized with a photograph

The problem

The heat order and the startup sheet define the heat — and they live in a folder until the end of the shift.

In a long steel mill, the heat order is printed from the MES in the morning and any change is written on it by hand: the real mix of scrap grades going into the basket, last-minute alloys, the melt shop lead's observations. The startup sheet is handwritten — by habit and by internal standard. Those two papers define the heat: chemistry, customer and destination grade. The metallurgical process is stable; the gap is in what happens to that paperwork afterwards:

• Late digitization — both documents can sit in a folder for four to eight hours before being digitized. In that window, what the paper on the floor says and what the system says do not match, and every decision taken downstream is taken without the startup data.
• Root cause that turns into guesswork — when a chemistry deviation appears downstream, there is no snapshot of the startup: nobody can check the real scrap mix or the alloys that actually went into the furnace. The five whys stop at the second why.
• An A3 with no data — the shift Kaizen A3 starts with no facts from the heat startup, so it almost always ends up fixing symptoms rather than causes. The continuous improvement engineer spends the Kaizen time reconstructing information instead of intervening.

Nobody sees that gap until an A3 has to be opened over a chemistry deviation or a customer claim, and the map of the heat has to be rebuilt from the annotated order and the startup sheet rescued from the folder. That is where a routine record turns into weeks of guesswork.

How it fits the IRIS system

Connect in photo mode — two photos at heat startup, and the heat exists in central memory at zero latency.

Heat startup does not need a new MES, nor a furnace lead typing at the furnace. It needs the information that already exists on paper — the annotated heat order and the handwritten startup sheet — to reach central memory at zero latency, without anyone transcribing it at the end of the shift. That is what Connect does in photo-over-paper mode.

The furnace lead photographs the heat order and the startup sheet with the industrial tablet or phone and carries on with the heat. Connect extracts the fields with an anchored LLM and lands them in the iLEAN central memory, available at zero latency to the metallurgist, to the rolling mill downstream and to the operational excellence team.

How Connect photo mode works on heat startup in a long steel EAF melt shop:

• A photo of the same paper as always — the furnace lead starts the heat exactly as before and takes two photos: the order printed from the MES with its handwritten notes and the handwritten startup sheet. Zero habit change, zero fields to type.
• Extraction anchored to the real templates — an LLM anchored to the plant's own forms extracts the structured fields: heat, basket, mix by scrap grade, alloys, target temperature, destination standard and customer. It does not generate free data; it recontextualizes what is already written, including the handwriting.
• Doubtful data held back — if a digit comes through ambiguous or a box does not match the template, the field is held and a person confirms it on the tablet before the record counts. Doubtful data never crosses on its own.
• Central memory at zero latency — validated fields land in the iLEAN central memory, timestamped and tied to the original image of each document: the startup snapshot exists from minute one of the heat.
• Live data for whoever needs it — the metallurgist checks the chemistry against the real mix, the rolling mill operator sees downstream which heat is coming and to which standard, and the operational excellence team opens the A3 with facts from the floor, not reconstructions.

See the full IRIS architecture →

Before and after

Paper heat startup vs. startup digitized with Connect

Aspect	Classic paper startup	With iLEAN Connect photo
Latency, order and startup sheet → central memory	4-8 h in a folder until the end of the shift	Zero latency — two photos and it is in
Root cause of a chemistry deviation	Guesswork: there is no startup snapshot to check against	Starts from the real scrap mix and the alloys, timestamped
Shift Kaizen A3	Starts with no data — fixes symptoms, not causes	Starts with facts from the floor, from minute one
Heat VSM	Estimated in the mapping workshop, from memory	Measured with live data from every heat
Continuous improvement engineer's work	Reconstructing information and chasing paper	Intervening on the process with live data
Furnace lead's habit	Same paper — plus somebody else transcribing late	Same paper — plus two photos with the tablet

Impact estimate

Impact estimate for your plant — to be validated with your own 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.

• Long steel mill with an EAF, several heats per shift, the heat order printed from the MES and annotated by hand, and a handwritten startup sheet required by internal standard.
• Connect photo pilot on heat startup — using the tablet or phone already on the floor, without changing the work standard or touching the metallurgical process. First value expected within a few weeks.
• Indicative payback of 4 to 9 months, depending on the number of heats per day. Estimate to be validated.
• Direct recovery of shift lead and operational excellence engineer time — the hours that go today into transcribing and reconstructing the startup return to the floor and to the Kaizen. Estimate to be validated against your data.
• The hard lever is that this case is the indispensable enabler for the rest of the iLEAN matrix: without the heat existing in central memory from second zero, no downstream analysis — chemistry, rolling, energy under ISO 50001, ResponsibleSteel reporting — starts from real data.

And the fair question from the lean manufacturing director

"What if the model misreads the furnace lead's handwriting or confuses an alloy on the startup sheet?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely recontextualizes a specific figure from one medium to another (reading the annotated heat order and the handwritten startup sheet, and extracting their fields against the plant's real template), the best models brought the error below 1.5% ^[1]. And even then, the critical call is not made alone: doubtful data is held and a person confirms it on the tablet before the record counts in central memory. The three safety rings exist precisely for this.

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