What is sovereign AI?
Sovereign AI is AI you own outright: the model weights, the index it retrieves from, the prompts, the traces and the record it builds all live inside a boundary you control, on hardware you can point at.
The test we use is simple. Can you draw every boundary your data crosses, and name who controls the far side of each one? If you cannot, it is a marketing position rather than an architecture.
What does air-gapped actually mean here?
It means there is no inbound path into the deployment at all. In a VarahiOne installation for a mid-sized business, nothing reaches in from outside.
Model updates come in signed, on your schedule, applied by your own admin. Outbound business traffic — the mail you send, the calls you place — leaves through the channels it always did. Nothing is copied out to be processed.
Is this the same as private cloud or a single-tenant SaaS?
No. In a private cloud your data still sits on infrastructure somebody else owns and operates, under their jurisdiction, and your access to it depends on a commercial relationship continuing.
Our deployments run on a machine inside your building. If you stopped working with us tomorrow, the record would stay where it is.
Can AI work without an internet connection?
Yes, and for industrial work it has to. Every VarahiEdge deployment is specified to run with the uplink pulled out, because in a workshop or on a plant floor the network is a maybe.
That constraint is what drives the design: models sized to run on a box you can buy and cool, rather than models that need a datacentre and a line you do not control.
Why use a small language model instead of a large one?
Because a workshop does not service every machine ever built. It services the dozen or two on its own floor, with a repair history belonging to those machines and no others.
A model built to know a little about millions of things spends nearly all of itself on things you will never touch — and that capacity is exactly why it needs a datacentre to live in. A model sized to your fleet fits on one ordinary computer, answers at reading speed, and is still there when the line drops.
Does our data leave the building to make the model better?
No. The signals, the job cards, the photographs and the verdicts your people give all stay where they were generated.
That record is available to improve the system precisely because it never had to be sent anywhere for the system to work in the first place.
How do you know a model is actually working?
We build an evaluation set with your domain experts before any training run — a written definition of what good means — and measure against it throughout. It is the only thing that decides whether a model ships.
Every model arrives with that evaluation set, its failure modes and its false-alarm rate. We would rather tell you a model is not ready than let you find out on the line.
What is a company's second brain?
A knowledge repository that is also a decision-making assistant: 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 one. So it is captured as a by-product of the work — a technician confirms a fault and that is a memory; an operator corrects a step and that is a memory; an engineer rejects a design and says why, and that is a memory.
Can a mid-sized business do this, or is it only for large enterprises?
A mid-sized business is the case we designed VarahiOne for. A large enterprise solves fragmented systems by hiring a platform team to integrate six vendors; a forty-person company cannot, and never will.
The answer is one core with everything in the same index, sized to a single on-premise server for a typical SME — not six integrations.
How is this different from the AI features our existing vendors added?
Each of those assistants is good inside its own slice. None of them can answer a question that crosses slices, and every question that matters crosses slices: can we commit to March, what did we promise this customer last time, why did this price change.
Answering those needs the mail, the drawing, the plan and the person at once, which means one index rather than five clever assistants.
Does the AI make decisions on its own?
No. Every Varahi system proposes; a person decides. The machine says what it thinks and why; a human says yes or no, and why; the verdict returns to the record, inside your own boundary.
Agents get a named scope, a spend limit and an audit trail, and they stop at the edge of it. An agent may open a conversation; a person closes a deal.
What hardware do we need?
For a typical small or mid-sized business, a single on-premise server sized to your headcount and the size of your corpus. Models are chosen to fit a box you can actually buy and cool.
For VarahiEdge, a box in the building where the machines are.
Are we locked in?
The opposite is the design goal. Every engagement ends with runbooks, failure modes, on-call paths and the second brain of the engagement handed to your team.
We are a dependency you can remove. If you stop working with us, the record stays in your building.
How does an engagement start, and what does it cost?
It starts with a conversation about a real constraint — the line, the asset or the process that is costing you, and what better would be measured in. The first conversation is with an engineer, not a salesperson.
Scope and duration are set with you once we have seen the data and the constraint, rather than quoted in advance. Each phase ends with a review you attend, and the evidence in front of you decides whether the next one starts.
Where is Varahi Technologies based, and who do you work with?
Bhugaon, Pune — near Chandani Chowk, on the road out to Pirangut. The company has been building software since 2011.
We work with industrial groups and mid-sized businesses across India, Europe and the USA. The sovereign AI deployment work is built for defence, pharma, BFSI, automotive and the public sector.
If the question you have is not here, it is probably the interesting one. The first conversation is with an engineer, not a salesperson — bring the constraint you are stuck on.
Talk to an engineer →