Defmate Is Building Sovereign AI For India’s Defence

Gurugram: Ask most people what sovereign AI should look like and they describe one giant model built entirely in India. For defence, that picture answers the wrong question. Information now reaches a headquarters faster than any staff can read it; the hard problem is deciding what to do with it, quickly, under authority, with a record that can be examined afterwards. Defmate, a defence technology company based in Gurugram, is building sovereign decision infrastructure around that gap: a family of small language models (SLMs), each sized to a single job, on top of a proprietary rule-based engine that decides which of them, if any, gets listened to.

The first generation of Indian security technology solved a collection problem. Fusion centres, monitoring systems and open-source intelligence platforms now bring scattered feeds into one place, and that work was necessary. The step after it remains open: once the picture is assembled, who decides what to do, on what rule, under whose authority, and how is that decision examined later. Defmate is building for that step.

A Dependency That Is No Longer Hypothetical

In July 2025, Microsoft suspended Outlook, Teams and other cloud services for Nayara Energy, India’s second-largest private refiner, citing EU sanctions tied to Rosneft’s stake in the company; access was restored within days after Nayara petitioned the Delhi High Court. Commenting on India’s wider technology dependence in its wake, a retired Army officer was widely quoted asking India to picture a harder scenario: “Now just think, it’s a hot war situation. Your entire ICT goes phuttt,” as reported by The Economic Times. His remarks addressed India’s dependence broadly, not Defmate. But the underlying risk, a foreign vendor able to suspend access at a time of its own government’s choosing, is the one Defmate Sovereign is built against: no vendor API, no external key, no outbound path, running entirely on the customer’s own network and hardware, air-gapped where required.

The Deterministic Core

That engine is called Nirnay, Defmate's own decision support system engine, designed and built in-house. A prediction from any model, Defmate's or anyone else's, enters Nirnay as an input, never as an instruction. A model can help someone understand a situation; a separate, fully deterministic system is the only thing that produces a verdict, against a rule set the deploying authority approves and a named officer signs: no network call, no file access, no SLM anywhere in that path. The company says the output is identical down to the byte whether or not a model is reachable, and that its test suite proves this by cutting the connection. Every rule that fires is shown to the officer before it runs, and every decision is sealed with two SHA-256 hashes into a tamper-evident audit chain, with no verdict-path failures recorded to date.

“There is no version of this where the SLM makes the call,” a Defmate spokesperson said. “A model can explain a decision. It cannot make one, change one or overturn one. That is the whole point of building it this way.”

The SLM itself has three jobs: understand what an officer is asking for, explain a sealed decision in plain language, and answer follow-up questions. If it invents a figure or shifts an outcome, the output is discarded before anyone sees it and the officer is told which check it failed. A companion staff assistant built on the same guardrails offers drafting and analysis skills (SITREPs, OPORDs, briefs, charts and decks) behind a privacy layer that runs locally in the officer’s own browser, on the same air-gapped network, screening detection rules across categories of sensitive information before any text reaches the model, and that has scored 94 out of 100 in blind quality tests against unmasked output.

Built for Every Echelon

The same engine runs under models scaled to where a decision is made: from a 4-billion-parameter model at 3.2 GB for a forward node with no signal, through 12- and 31-billion-parameter tiers for a squadron workstation and a site server, up to a 70-billion-parameter tier for a command post, with a frontier tier above 120 billion parameters in development. The rule set, engine and seal are identical at every size; only the model doing the talking changes, and the whole stack is designed to run fully offline for weeks if required.

“We wanted the smallest model that could reliably do one job, deployed where that job is done,” the spokesperson said. “The decision layer beneath every tier has to be the same one, or you do not have a platform, you have five products.”

Defmate applies four tests to the word sovereign, and holds that all four must pass at once: the data stays inside Indian systems; the system runs on Indian infrastructure, including with no connection at all; India owns the code and the rules; and a named human can see the reasoning and overrule it. Defmate's position is that partial measures against these tests do not add up to sovereignty. “If something is sensitive it should never leave the device,” the spokesperson said. “Redacting it after it has gone out is a different promise from the one we are making.”

The World Model Bet

The more ambitious part of Defmate’s work sits further out, and the company is deliberate about how it describes it. A language model learns which words follow other words. A world model learns which situations follow other situations when someone acts: given the state of things right now and an action taken inside it, predict what happens next. For defence, that is the more useful problem, because the real question is rarely what happened. It is what happens if a particular decision gets made.

“We are working toward a world model, and we are being disciplined about how we get there,” the spokesperson said. “What we have is a bet, a bounded starting point where we can measure whether the bet is paying off, and prediction work already in use that the bet grows out of.”

Defmate’s path starts narrow: bounded operational environments where state, actions and outcomes can all be measured, and where its predictive work already produces forecasts with a stated confidence. The next step is counterfactual (given the current state, what does each available action lead to over the coming days), requiring a forward model over a small, well-defined state space, trained on data only the customer holds. From there, the roadmap runs through simulation and digital twins, and only then toward the action-conditioned prediction a genuine world model requires.

Nirnay tests any prediction against doctrine, standard operating procedures, operational constraints and the authority of the officer on duty, and it permits, conditions, escalates or refuses the action, escalating whenever it is not confident. “Everyone is building AI that decides,” the spokesperson said. “We are building the layer that decides whether the AI is allowed to. The models will keep changing, and most of them will never be Indian. The doctrine layer is the part that has to be ours.”

Every decision the system touches becomes a record of what an operator saw, what the system recommended, what the operator actually did, and what happened next, kept inside the customer’s own boundary. “We want to ship something small and unglamorous that works,” the spokesperson said. “The bigger model is what you earn.”

The company also points to where policy is heading. Atmanirbhar Bharat and Make in India set the direction; India’s draft Defence Acquisition Procedure 2026 sharpens it, describing ownership of software source code and core architecture as crucial to long-term strategic advantage, and proposing a move toward self-certification with third-party audit. Should that audit requirement be adopted, Defmate says, any AI supplier to Indian defence would need an evidence record of how its system decided, and Nirnay’s signed, replayable audit chain is built to serve as one.

About Defmate

Defmate is a defence technology company building sovereign decision infrastructure for the Indian armed forces, founded by alumni and professors of the Indian Institutes of Technology (IIT). Its core engine, Nirnay, is designed and built in-house and runs on-premises and air-gapped where required, with the technology proprietary to Defmate and owned in India, and its models built, fine-tuned and retrained in India. The company says the product is shaped in ongoing collaboration with serving and retired defence personnel, whose operational input into requirements, workflows and evaluation guides what Defmate builds next.

More at https://defmate.com/

###