A plant head sits through a demo of a diagnostics tool his OEM is weighing. It is sharper than he expected — it already recognises faults his own team only recently learned to name.
He asks how it got that good. The salesman is proud of the answer. The model has learned from fault logs sent up by plants like his, from all over the world, into an account none of those plants can see inside. His logs, if he signs today, join the same pile. The model gets a little sharper. He does not.
He asks what comes back to him for that. A support dashboard. A ticket queue. Not the record itself — that stays wherever the vendor keeps it, built out of work his own technicians did on his own floor, confirming faults the manual never named. He leaves without signing.
Why we build it this way
The asset in that room was never the model. It was the knowledge the plant had already produced, one confirmed fault and one corrected procedure at a time, sitting in logs nobody outside the plant had ever thought to call valuable — until a vendor worked out how to collect them.
The obvious answer in this industry is to use whichever model is sharpest, wherever it was trained or hosted. For most software that is a fine rule. For this it is the wrong one, because the thing a plant pays years to learn — how its own machines fail, what its own people do about it — is exactly what gets folded into someone else's product the moment it leaves the building. The plant gets an answer back. It does not get the knowledge back, because the knowledge is now the model's, sold to the plant down the road too.
For an Indian plant this cuts a particular way. Most of what makes an Indian plant run — the local parts substituted in, the power that comes and goes, the maintenance habits built around both — is knowledge nobody writes a manual for, because nobody outside the plant was ever in a position to ask. If that knowledge trains a model somebody else owns, the expertise of running a plant in India moves with it.
The knowledge of how an Indian plant runs should stay in India, owned by the people who generated it.
That is not a preference about where a company registers its office. It is a belief about where the knowledge should end up, and the only way we know to keep that promise is to build the record so it never has to leave in the first place.
How it works, roughly
The record a plant builds with us lives on hardware inside its own boundary — a box the plant holds, not an account at a vendor somewhere else. What it reads:
- the fault a technician confirmed, in the words he used for it
- the procedure a supervisor corrected, and what was wrong with the version before it
- the local workaround — a substituted part, an adjusted tolerance — that nobody outside the plant would think to ask about
What it gives back is the plant's own history, read to whoever is asking, rather than an answer borrowed from plants it has never met. When a person disagrees with what it proposes, the correction becomes the next thing the record holds — not a signal sent elsewhere to sharpen someone else's version of the same tool.
A year on, the plant knows more about itself than it did a year before, and every bit of that knowing stays where it was earned. Nobody down the road got sharper because of it. A person still reads what the record says and decides what to do, and the record, like the decision, never left home.
Read the vision →