Windows NT screen in pharma packaging — read the HMI without touching the machine.

The problem

The machine does its job. Its screen tells nobody about it.

In pharma packaging it is common to find lines qualified 15 or 20 years ago whose HMI runs on Windows NT (or 2000, or XP), with no updates, no network, and a proprietary driver the manufacturer no longer maintains. The machine packages perfectly. The problem is not the machine — it is that its screen is the only way to know what is happening, and nobody outside the line ever sees it.

1. It cannot go on the network. Connecting that HMI to the MES or the corporate network means exposing an unpatched Windows NT to the internet — no OT manager signs that off.
2. It cannot be replaced. Changing the HMI means requalifying the whole line. Cost and time nobody wants to take on while the machine works.
3. That is why counts are written down by hand. Every hour, every batch change, the operator reads the screen and fills in a sheet. At 10pm somebody types it into a spreadsheet. The next morning, somebody else moves it into the corporate system. By the time it reaches management, it is already yesterday.

The data exists, it is on the screen, it is exact. It just lives on an island — and yet that count is what production needs to plan, quality needs to review and management needs to decide. Gap number one from the book: data islands.

How it fits the IRIS system

iLEAN Connect — the filler between the legacy screen and the Annex 11 ready system.

The IRIS system defines three capture levels for machines depending on their state: manual (photo of the panel), intermediate (reading the local computer even if it is old) and integrated (modern interface). Here we are exactly in the hard intermediate case: there is a local computer, but it cannot be touched. Connect solves this scenario without asking you to change the machine — the filler that covers the dead zone your HMI left exposed.

A camera looking at the HMI. OCR + visual template. Data into the Annex 11 ready system. The machine is not touched. The OT network is not touched. The qualification is not broken.

How Connect works applied to the Windows NT HMI in pharma packaging:

• Capture by vision, not by network. An industrial camera points at the HMI screen. An iLEAN terminal nearby carries the OCR + visual template model trained for that specific view. It captures every few seconds. Without touching the machine, without opening its OT network.
• HMI view discrimination. Connect can tell apart the normal operating view, the alarm view, the menu view and the recipe view. It only extracts counts when the valid view is on screen — if the HMI goes into alarm, it logs it as an event, not as a wrong count.
• Push to the Annex 11 ready system. The data (count, batch, format, alarm) reaches the plant's receiving system with timestamp, hash of the original image, extraction model and full traceability. The original image is stored as evidence.
• It works with no network. If the plant loses WiFi, the Connect terminal keeps capturing and buffering locally. When the network comes back, it syncs. What is critical does not depend on connectivity.

Connect carries the data, the agent structures it, the person validates in the receiving system. The GMP electronic signature is made by QA, not by Connect. That is the rule — always.

See the full IRIS architecture →

Before and after

Manual counting off the HMI vs. vision capture with Connect.

Aspect	Operator writes the screen down by hand	With iLEAN Connect (vision on the HMI)
Counting frequency	Every hour or every batch change	Every few seconds, continuous
Latency to the system	Hours (sheet → spreadsheet → system)	Seconds to the Annex 11 ready system
Changes to the machine	None (because of qualification)	None — external camera only
Reading error risk	Whatever any tired human makes	OCR anchored to a template, with the image as evidence
GMP traceability	Paper sheet + spreadsheet transcription	Hash + timestamp + image + extraction model
Cost of keeping the legacy line	The same — and you are still blind	The same — and now the screen speaks

Impact estimate

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

The block below is an estimate to be validated with the specific data of your line. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Pharma packaging line with a Windows NT (or 2000/XP) HMI isolated from the network, counts written down by hand each shift and pushed to the corporate system hours later.
• Connect pilot on one line (camera + terminal + integration with the Annex 11 ready receiving system). First value expected in a few weeks: counts reach the system in seconds, not in hours.
• Indicative payback between 4 and 9 months, depending on operator hours spent on transcription, rework from mismatching counts and the cost of planning blind.
• Hard lever: ≥ 30% reduction in operator time spent writing the screen down — and the end of "it is already yesterday by the time the data goes up". Planning is done with data from the current shift, not the previous one.

And the OT manager's reasonable doubt

"What if the OCR gets it wrong and puts a bad count into the GMP system?" — hallucination is a problem of free generation, not of anchored tasks. Reading figures off a screen with a known visual template is exactly an anchored task, where the best models brought the error below 1.5% ^[1]. And even so, the original image is stored as evidence for the auditor, and the receiving system is still the one that qualifies and signs. The three rings of safety are there precisely for this.

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