Two models, never one: the machine understood by science, and the person accountable for it understood by memory. Physics written down before any training run; decisions and reasoning recorded as the work happens.
An engineering company in Pune, working with industrial groups and mid-sized businesses across India, Europe and the USA — and recruiting the next generation of leaders and builders.
Two product stacks and one services practice. The products share a runtime and an evaluation discipline; the services are bought separately as engagements and run on whatever stack you already have — no product purchase required. The fourth card is the thinking underneath all three.
An engineering company in Pune that takes on the hard technologies and finishes them. Who we are, how we work, the second brain we are building, and what it takes to become a Varahian.
For small and mid-sized businesses: Varahi Workspace for email and drive, Happen for CRM, vision and multi-agent outreach, VarahiConverse for projects and HRMS — all on one air-gapped native AI core.
Four modules on one edge core: VEDA reads machines with small models, Nirdesh replaces the OEM manual with voice, Drishti watches quality and posture, VCAD generates parts with structure and carbon.
Sovereign AI architecture and agent ecosystems, delivered as engineering rather than advice. Bought individually, run on your own stack with your own models — no product purchase required.
VarahiOne is our stack for small and mid-sized businesses: Workspace, Happen and VarahiConverse on one air-gapped native AI core. Around it sits the wider product portfolio, built agent-first from the first commit.
A product earns a name here only when it is sold separately. Everything else stays a feature with a plain label.
The sovereign runtime: model serving, retrieval, agent scheduling, budgets and an append-only audit log — installed inside the customer boundary. Everything else in the ecosystem runs on it.
The whole company on one air-gapped native AI core: Varahi Workspace for email and drive, Happen for CRM, vision and multi-agent outreach, VarahiConverse for projects and HRMS.
Screening workflows, from intake to verdict, with the audit trail a regulated process needs.
Voice and call intelligence: transcribe, structure and route conversations without shipping audio off-site.
Investigation and retrieval over an organisation's own corpus — provenance attached to every finding.
Industrial data pipelines with models in the path — from tag to decision, instrumented end to end.
Agents are not products. They get a descriptor on first use, never a logo, and they run under a named scope inside VarahiOS.
Reads traces and history, proposes a fault tree, defers to the technician. Learns only from confirmed verdicts.
Handles inbound requests end to end, hands over to a person the moment the scope runs out.
This division puts models where the metal is: on the line, in the vehicle, on the plant floor — perception, control and diagnostics running on hardware you already own.
Constraint we design to first: the network is a maybe. Every deployment in this division is specified to run with the uplink pulled out.
Small language models trained on the twelve machines you actually own — reading bus signals and service history to propose a ranked fault tree the technician confirms or rejects.
The next generation of the machine manual: ask out loud, get walked through the procedure a step at a time. Gloves on, no screen, no signal. Corrections stay for everyone.
One vision core looking at two things: whether the part is in tolerance, and whether anyone is working in a way that will injure them. Bad stations produce both — measured as joint angles, never faces.
Describe the bracket, get editable geometry — with the load path traced through it and the embodied carbon broken down by material and process, at quotation time rather than in a report a year later.
One box in your building. One record of what the plant knows. Every module proposes; a person decides.
Explore VarahiField in detail →No slideware transformation programme. An architecture, a running system inside your boundary, and the evidence that it works.
The evaluation set decides, not the calendar. If it stops moving, we say so rather than spending your budget to a deadline.
Air-gapped or isolated inside your own tenancy. Model weights, vector store, prompts, traces and evaluations all live inside the boundary you control. Every boundary the data crosses is labelled on the drawing, because unlabelled boundaries are how sovereignty claims get lost.
Not one chatbot. A named set of agents, each with a scope, a budget and an audit trail — the same discipline we run on our own codebase: propose-only, budgeted, daily, with a human approving every merge.
Pluggable adapters, so the same agents run on whichever coding CLI or model your team already uses. Swapping the model should be a config change, not a project.
Drawings, SOPs, service records, tickets and tribal knowledge, made retrievable with provenance on every answer. An answer without a source is not an answer.
We build the evaluation set with your domain experts before any training run, then measure against it throughout. It is the only thing that decides whether a model ships.
Runbooks, failure modes, on-call paths and the second brain of the engagement, handed to your team. We are a dependency you can remove.
Read it top to bottom: your systems feed in, our layers do the work, a person signs off, and the verdict returns to your record. Everything below the dashed line lives inside the site boundary. Blue is ours, gray is yours, orange is where a person decides.
Scope and duration are set with you once we have seen the data and the constraint — not quoted in advance.
Three cores on one foundation — VarahiOne for the business, VarahiField for the shopfloor, AI Transformation for the boundary — all wired to a second brain in the middle, with a person deciding above it.
Most companies buy AI from a slide. These are the actual architectures — one per thing we sell, plus the way we think about all of it. Turn each one over, then read where the orange is: that is where a person still decides.
Every AI company says it is building intelligence. We think that is the wrong noun. What a plant, a workshop or a mid-sized business is actually missing is not intelligence — it is memory: the ability to know what it already knows, and to bring that knowledge to the person who needs it, at the machine, at the moment they need it.
So we build two models, never one, and we keep them separate on purpose. One is a model of the machine. The other is a model of the person accountable for it. They are built by different disciplines, they fail in different ways, and a system that blurs them ends up trusting a statistic where it needed a physicist, or a physicist where it needed a veteran.
A machine is not a mystery. It obeys conservation laws, it has tolerances, it runs to a takt time. We write those down before any training run, so the model is bounded by how the process actually works instead of guessing from whatever data happened to be lying around.
This is also why we are suspicious of our own results. A number without the method that produced it is a claim, not a finding. Every model we ship arrives with its evaluation set, its failure modes and its false-alarm rate — and we would rather tell you a model is not ready than let you find out on the line.
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 — not because it is secret, but because writing it down has always meant forms, and nobody fills in forms. So we build the writing-down into the work. A technician confirms a diagnosis: that is a memory. An operator corrects a step: a memory. An engineer rejects a design, and says why: a memory.
Over time that record becomes what we call the company's second brain — and it belongs to the company. Agents read it; people own it; it is handed over at the end of every engagement. A second brain on somebody else's servers is not a second brain. It is a dependency.
Instrument the process before modelling it. Anything that cannot be measured is not in scope — however interesting it looks.
The machine's model and the person's model must use the same words for the same fault, or neither can correct the other.
The machine proposes; a human says yes or no, and why; the answer returns to memory. That loop is the whole system, and the orange in every structure we draw is the promise we will not close it without you.
We are not trying to replace the person who knows the machine. We are trying to make sure the company still knows what she knew, after she has gone home.
The company's memory across email, deals, projects and people — on one core it owns.
The machine measured, the operator remembered — four modules that write the plant into its own record.
The same thesis, built on your own stack — with the boundary drawn so you can hand it to an auditor.