Reading the AR chamber’s panel without touching the machine

The problem

The asset that decides the finish is the only one leaving no trace.

An anti-reflective recipe is several layers of nanometric thickness, with its vacuum curve, its pumping times, its ion gun and its temperature. The panel shows it all, cycle by cycle — and overwrites it. The result is a hole exactly at the point where whether the lens comes out right is decided:

• Zero correlation between load and defect — when poor adhesion, an off-pattern reflection color or a halo appears, nobody can check what the chamber did in that specific load. The data existed during the cycle and is no longer anywhere.
• Root cause analysis depends on one person — it is solved by the veteran technician's intuition, which is usually right. The problem is not their judgment: it is that that judgment cannot be cloned when the workforce grows fast, and it leaves with them when they leave.
• The alternative is a CAPEX nobody signs — replacing the equipment costs what it costs and stops the coating line for weeks. The project is postponed year after year, and the data hole stays.

So the plant lives with a black box in its most delicate process, and every coating defect investigation starts with the variable that would explain the most — simply absent.

How it fits the IRIS system

Connect in photo mode on the panel — looking at the screen instead of integrating the machine.

The missing information is already on the screen: the panel shows it load after load. What does not exist is the path from that glass to the central system. Connect solves it the cheap way — by looking — without opening the machine's network, without touching its software and without voiding the manufacturer's warranty.

A fixed industrial camera in front of the screen, capturing every few seconds. OCR with an anchored model interprets each frame, the load's complete time curve is reconstructed and cross-referenced with the lab system to know which lenses were in the domes. Everything stays in central memory, queryable by batch.

How Connect operates on the AR chamber's panel:

• External camera, zero intervention — a mount is installed in front of the panel and that is where the intervention on the equipment ends. The cabinet is not opened, the network is not tapped, the machine's software is not touched. For the vacuum chamber, nothing has happened — and the manufacturer's warranty stays intact.
• OCR anchored to that specific panel — the model knows that screen's real layout: where the vacuum curve is, where the pumping times, where the ion gun and where the temperature. It reads values at known positions, it does not guess text.
• The complete curve, not snapshots — the capture every few seconds is reconstructed as the whole load's time series, from pumping to venting. That is what makes a drift visible, which is exactly what a loose photo does not show.
• Cross-reference with the lenses inside — the curve is associated with the lab system to know which lenses occupied the domes on that load. Without that cross-reference the curve is a pretty chart; with it, it is traceability.
• Drift detected before the defect — with several loads accumulated, progressive deviations start to show — a pumping that takes a bit longer each week — announcing the problem before the first lens comes out wrong.

See the full IRIS architecture →

Before and after

Vacuum chamber as a black box vs. chamber digitized with Connect

Aspect	Isolated proprietary panel	With iLEAN Connect on the panel
Load ↔ coating defect correlation	Nonexistent	Root cause analysis per batch, in minutes
The load's curve (vacuum, pumping, ions, temperature)	Shown and overwritten	A complete time series, queryable by batch
Traceability of which lenses were in the domes	Reconstructed, if possible	Automatically cross-referenced with the lab system
Detection of equipment drift	When there is already a defect	Before the defect appears
The veteran technician's knowledge	Lives in one person and leaves with them	Backed by data anyone can query
Equipment replacement	Pending, undated, on the table	Deferrable indefinitely without going blind

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.

• Optical plant with a 10-to-25-year-old anti-reflective vacuum chamber, an isolated proprietary panel and coating defects whose cause cannot be attributed to the load today.
• Connect pilot on the panel: fixed industrial camera, model anchoring to that specific screen and cross-reference with the lab system. First value expected within a few weeks.
• Indicative payback between 5 and 10 months, adding the replacement CAPEX not executed and the coating scrap avoided. Estimate to be validated.
• The deferred CAPEX is usually the figure that moves the committee: the replacement project is postponed without giving up the data that justified it.
• The underlying lever is closing the hole in the most delicate process: from here on, every coating defect has something to correlate against, and the veteran technician's judgment stops being the only record that exists.

And the fair question from the coating manager

"What if the camera misreads a value from a twenty-year-old panel and we make decisions on false data?" — hallucination is a problem of free generation, not of anchored tasks. Here the AI reads a numeric value at a known position on a known panel: in that class of task the best models brought the error below 1.5% ^[1]. On top of that the reading is redundant over time — many frames per load —, so a discordant reading is discarded against its neighbors instead of propagating, and when a frame does not reach the confidence threshold it is flagged instead of a value being forced.

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