A vendor pitches a workshop on an AI system that has read every manual, every forum, every fault, in every language there is. It sounds like more than any single shop could need. Then someone asks what happens when the shop's internet drops, as it does most weeks, and the answer is that the system stops answering until the line comes back.
That is the trade nobody puts in the pitch deck. A model built to know a little about everything has to live somewhere with the power and the connection to carry all of it — a datacentre, reached over a line the workshop does not control. A model built to know a lot about the machines a workshop actually owns can live on a box in that workshop, because it was never asked to hold the rest of the world's knowledge as well.
What "more" is actually for
A workshop does not have every machine ever built. It has the models on its own floor — a dozen, perhaps two — and a service history that belongs to those machines and no others. A model sized to carry general knowledge about millions of things spends most of its attention on things this workshop will never see. That is not a free choice: it is the reason such a model needs a datacentre to run in and a line to reach it.
VEDA is sized the other way. It is trained on the machines a given workshop or OEM actually services — the signals those units put on the bus, the repair history those units have built up, the notes the engineers on that floor have written. It knows less than a general model about the world, and a great deal more about this fleet.
What that buys back
The first thing it buys back is speed. A model this size answers as fast as it can be read, because the question never has to travel to a server and back. The second is availability: VEDA runs on the box in the corner, so a dropped line is a Tuesday problem, not a diagnosis problem.
The third matters most to an OEM watching over many workshops rather than one: the fleet's history never has to leave the building to be useful. The bus traffic, the job cards, the verdict a technician gives when they confirm or reject a suggestion — all of it stays where it was generated. It is available to make VEDA better precisely because none of it had to be sent anywhere to make VEDA work in the first place.
A model that has to know everything cannot live where the work happens. One built to know your fleet can.
The size question, answered properly
Small, here, is not a compromise made for cost or for speed, though it happens to deliver both. It is the size the job calls for: reasoning about a bounded set of machines a workshop or OEM already owns, using a record only that company holds. A bigger model would not know those machines any better. It would only be harder to keep next to them.
VEDA sits under VarahiEdge, our OEM engineering AI division, alongside Nirdesh, which answers from the machine manual, and VCAD, which turns a description into a part. All three run on the same on-premise box, inside the same building as the fleet they serve.
Read about VarahiEdge →