The IPT lead’s earpiece — operational knowledge enters the system in real time

The problem

Critical operational knowledge lives in the IPT lead's head and does not enter the system in real time.

At a hormonal manufacturing site, the Production Lead is not sitting in front of a screen: they are walking between containment suites, hands busy, wrapped in PPE, seven hours a day. They are the person who first sees when something starts to deviate:

• An incipient deviation in granulation — a batch starting to behave differently long before the data reaches any batch record review.
• A change in a suite's operations — an adjustment, a handover, a replaced consumable — that today nobody notes in the moment because inside the PPE there is nowhere to type.
• An observation that should be a deviation — and gets reconstructed hours later, from memory, if it gets reconstructed at all.

At a site with an FDA audit on the radar, that undocumented information is not only an inefficiency: it is warning letter risk. What is not written, for the auditor did not happen — or worse, happened and nobody documented it. And the site's continuous improvement plans — Hoshin Kanri, Lean Six Sigma, the MPS — depend on that knowledge existing as data: today they feed on what someone manages to remember at the weekly meeting.

How it fits the IRIS system

Connect in hands-free voice mode — the IPT lead takes notes by dictating, without letting go of anything.

The earpiece is not a system recording the IPT lead during the shift: it is a tool they activate voluntarily when they want to note something. They decide when to speak and what to dictate. Connect structures it, ties it to the plant's active context and returns it in seconds.

The IPT lead says "iLEAN, note this". They dictate the observation in their own words, walking between suites. Connect structures it as a preliminary deviation incident tied to the active batch, line and shift. Seconds later, the return arrives by voice on the earpiece or as a push to the tablet.

How Connect operates in earpiece mode at a hormonal pharma site:

• A personalized wake-word — the IPT lead opens the dictation with their own phrase pronounced deliberately. Outside that phrase, the earpiece transcribes and sends nothing.
• Free dictation, no forms — they speak in their own words, just as they would tell a colleague: "the suite 3 granulator has been giving odd granulometry since the batch change".
• The LLM structures the incident — Connect turns the dictation into a preliminary deviation incident with clear fields, automatically tied to the batch, line and shift active at that moment, without the IPT lead reciting those data.
• A return by voice or push — the confirmation arrives on the same earpiece or, if more context is needed, as a push notice to the nearest tablet. The IPT lead keeps walking, with the observation already in the system.
• Human verification before signing — nothing signs as a formal deviation without passing through the early human verification on the tablet: a person reviews the draft, completes it and escalates it per the site's SOP.

See the full IRIS architecture →

Before and after

Knowledge that dies at the end of the shift vs. knowledge that enters the system in real time

Aspect	Knowledge dying at shift end	Knowledge entering in real time
An observation caught between suites	Stays in the head until the end of the shift, if at all	Dictated in the moment, hands-free and with the PPE on
The link to batch, line and shift	Depends on someone remembering and reconstructing it	Automatic, the system knows what is active
The record as a deviation	Written hours later, from memory, or not at all	Enters as a preliminary deviation in real time, pending human verification
Cross-reference with the MBR's data	Impossible: the data does not yet exist when it happens	Crossed with the MBR's live data from the first minute
Preparation for an FDA audit	Undocumented operational knowledge, a potential finding	Observations traceable from the moment they happen
Continuous improvement (Hoshin Kanri, Lean Six Sigma, MPS)	Fed by recollections at the weekly meeting	Fed by real data captured in the plant

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.

• Hormonal manufacturing site with a Production Lead walking most of the day between containment suites, hands busy and with no practical keyboard access inside the PPE.
• Connect earpiece pilot on one shift or one line — without changing the IPT lead's role, without adding administrative tasks.
• The return is not a countable hour saving: it is enabling Hoshin Kanri, Lean Six Sigma and the MPS with real data instead of memory reconstructions at the weekly meeting — and reducing documentary exposure at an FDA audit.
• The impact estimate in deviations captured in time and in the quality of the continuous improvement plans is an estimate to be validated with the site's continuous improvement manager, not a generic figure invented here.

And the fair question from the plant manager

"Doesn't this end up being a system recording everything my Production Lead says?" — no. It is exactly the other way round: the earpiece listens to nothing until the IPT lead activates it with their wake-word. There is no background recording, no continuous transcription, nobody listening to their conversation with a suite operator. The only thing entering the system is what they decide to dictate — because they want it recorded, not because someone above asked. And the second doubt, "what if the AI invents half a deviation?": for anchored tasks like this — structuring a dictation against the active batch, line and shift — the AI's reliability is very far from free generation's hallucination problem ^[1], and even so nothing signs as a formal deviation without passing through the early human verification on the tablet.

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