A model of the machine, and a model of the person who is accountable for it. The first is a matter of science: a plant has conservation laws, tolerances and a takt time, and those get written down before any training run so the model is constrained by the process rather than guessing at it.
The second is a matter of memory. The most valuable knowledge in any plant is in the head of the person who has run it longest, and it has never been written down because nobody had a way to do it that did not mean forms. We build the two separately, on purpose, and we make them share one representation — the same words for the same fault.
This is where every other idea on this page comes from. Get the thesis wrong and sovereignty is just hosting, agents are just automation, and a second brain is just a document store.
Every architecture we publish has the same three things: something measured, something remembered, and exactly one orange element — where a person decides.
That convention is not decoration. It is the thesis, drawn.
A confirmed diagnosis, a corrected procedure, a measured deviation, a rejected design option — each is small. Written into one record, over three years, they become the thing a competitor cannot buy.
That is why every product and every engagement writes into the same place.
Our objective — across every product and every engagement — is to build the customer a second brain: a knowledge repository that is also a decision-making assistant. Not a chatbot on top of a document store. The company's own reasoning, recorded as it happens, and usable later by someone who was not there.
The trick is that nobody will fill in a form to build it. So it is built as a by-product of the work: the technician confirms a fault and that is a memory; the operator corrects step four and that is a memory; the engineer rejects a design and the reason is a memory. The tools are useful on day one, and the brain is what accumulates.
The three offerings are instruments of this one objective. VarahiOne writes the business into it. VarahiEdge writes the shopfloor into it. AI Transformation builds it for a company that already has its own stack.
If the objective is a company that knows what it knows, the knowledge has to live where the company can see it, keep it, and take it with them. That single requirement decides the architecture: models that run on hardware you own, inside a boundary you can draw, working when the internet does not.
It also decides who we build for. A shopfloor with no signal in the shed. A defence supplier that cannot send a drawing to a cloud. A mid-sized business that has watched five vendors each take a slice of its data. Sovereignty is not a control we add. It is the condition under which the objective is even possible.
And it is why an Indian engineering company should be the one doing this. The knowledge of how Indian plants actually run should stay in India — measured, remembered, and owned by the people who generated it.
Can you draw every boundary your data crosses, and name who controls the far side of each one?
If not, it is a marketing position, not an architecture.
Four words carry the whole thesis. Build — we are engineers, and the thing gets made. Minds — a second brain for the company: the machine understood, the people remembered. You own — it runs on your hardware, inside your boundary, and it leaves with you. Here is a day in which all of that is true.
Up to here the path is common: two models, one record, inside your own boundary. It diverges at the last step — an industrial group runs it across plants it already owns, and a forty-person business builds new things on top of it that it could never have staffed. Pick the one you live in.
A plant inside a group that has run for forty years, and has finally written down what it knows. The day is about a company seeing itself.
A machine starts with a sound it did not make yesterday. The man who would once have recognised it retired in 2027. The plant recognises it anyway, because he spent his last fourteen months confirming what the system proposed, and every confirmation became memory. It proposes a cause. The shift lead agrees. That agreement is a memory too.
A customer asks whether March is possible. The answer used to take three people a morning. Now one person asks, and the company answers from its own record — the email where the promise was made, the drawing revision it referred to, the person on leave who is the actual constraint. She commits to April, with the reason, in eleven minutes.
A fixture drifts by less than a millimetre. Nobody sees it, but the station does — and it notices in the same breath that the operator has been reaching further to compensate. Both are flagged together, because they are the same problem. The station is rebuilt that afternoon, not in six weeks when the batch comes back.
A European buyer asks for the embodied carbon of a bracket before ordering. The design engineer does not start a project. The number is already on the part, broken down by material and process, because the part was generated with it. The quote goes out with a figure the company can defend.
A customer's security team wants to know where the company's AI sends its data. The answer is a drawing with one gate on it, and a log of everything that has ever crossed. The meeting is short. Nothing on the drawing lives outside the building.
Every one of those moments ended with a person deciding. The machine proposed; a human said yes, or no, and why. That is the one thing we will not automate, and the orange element in every structure we have ever drawn is a promise that we will not.
A forty-person business on one box in a cupboard, running VarahiOne — Workspace, Happen, VarahiConverse on a single air-gapped core. The day is about a company small enough to change on Tuesday, building things on the stack it could never have staffed.
There is no data centre and no platform team. There is one server the founder bought, and the whole company runs on it — mail, drive, pipeline, projects, people. No line item for a seat that grew tenfold, no vendor holding a slice of the record. The IT department is a person who also does accounts.
The pipeline says nine live deals, and it is right — because nobody typed it. It was read out of the mail, the calls and the WhatsApp thread where the big one actually started. The founder stops running the company from memory and a spreadsheet, and starts running it from a record everyone can see.
A production planner who is not a programmer builds an agent on top of the stack — one that reads incoming drawings, checks them against every job the company has ever quoted, and flags the ones that will go wrong on the machine they own. It took an afternoon, because the record was already there and the agent only had to be pointed at it. In 2025 that was a funded project. Now it is a Tuesday.
An enquiry arrives that is three times their usual size, with a compliance annexe that would once have cost a week they did not have. The company answers it out of its own history — the jobs that resemble it, the certifications, the two clauses it should not accept. They bid. Being small stops meaning being shut out.
In a company of forty, one resignation used to take a quarter of what the company knew with it. This one does not. The reasoning is in the record — why that supplier, why that tolerance, what was tried in 2028 and did not work — and the person who replaces her is useful in a fortnight instead of a year.
Same promise, at a different scale. Every one of those moments ended with a person deciding — and the whole record sits on a box the company owns, so if they ever stop working with us, the second brain stays in the building. That is the orange element, and it is the same on both paths.
The value was never in the AI. It was in a company finally being able to see itself — and owning what it sees.
A machine, a line, a process, or a question your company cannot answer about itself. First conversation is with an engineer.