Ten use cases, by function
1 · Maintenance and fault-finding
Technicians describe a fault in plain language and get the right procedure, history and likely spares from the manuals and records, offline, on a tablet.
2 · Logistics and sustainment
Agents reconcile stock, forecast consumption from usage records and draft indents for a logistics officer to approve.
3 · Manuals and SOPs, answered
Questions answered only from approved technical publications and standing orders, with the paragraph cited, in English or Hindi.
4 · Report summarisation with redaction
Long reports condensed for commanders, with personnel, unit and location identifiers irreversibly redacted by Veil before any model reads them.
5 · Imagery analysis support
Detection on imagery flags what an analyst should look at first. The analyst decides. This is the work of our Astra line, engaged with design partners.
6 · Cyber alert triage
Agents enrich and group security alerts, summarise the likely story and propose containment steps for the analyst on shift.
7 · Secure software engineering
Coding, testing and review assistance for mission software teams, with every change reviewed and merged by a person. Our coding family, Taksha, is generally available.
8 · Procurement files
Tender documents compared against requirements, gaps and compliance issues listed, and a draft evaluation note prepared for the committee.
9 · Training content in Indian languages
Course material, question banks and briefings drafted from approved sources, in the languages trainees use, for instructors to review.
10 · Disaster relief coordination
When civilian networks are down, agents track requests, supplies and observe-only drone sweeps, and keep the coordination picture current for the officer in charge.
The line we do not cross
No targeting. No weapons release. No autonomous engagement. We state it first because it shapes everything else. Our agents and models are not offered for fire-control, targeting or any weapons chain, and they are never placed in a safety path or a control loop.
That is not only an ethical position; it is an engineering one. The systems that take lethal decisions need certification regimes, command chains and accountability that no general-purpose AI platform should pretend to provide. The large, urgent opportunity is everywhere else.
The three non-negotiables
Inside the wire, accountable
No link out. The platform deploys air-gapped with Swaraj, on hardware you own, with no dependency on an external service to keep working.
Lineage you can account for. Every model is registered with its provenance, evaluation history and approver in Model Hub. For the edge, Tatva Edge and Tatva Nano are trained from scratch in India.
A human decides, and it is recorded
Named approvers. Every consequential step waits for the authority it belongs to, with escalation if they are unavailable.
A record that survives an inquiry. Every call writes a signed Trust Receipt: what was asked, what the model saw in redacted form, which model and version, who approved and when, verifiable without trusting the operator.
Where to start
Start with a document-heavy process that a person already reviews: technical publications, maintenance records, or report summarisation with redaction. These deliver value quickly, carry low risk, and build the approval and record-keeping habits that more sensitive uses will need.
We brief defence and security teams in person, with demos on disconnected hardware. Request a closed-door briefing. For the edge, read Tatva Edge for defence; for the governance model, read the sovereignty test.
What we don't claim
We do not claim accreditation for any specific classified network, integration with any specific service's systems, or that an AI system is correct because it is governed. Integration is done with design partners, under the agreements a programme requires. Astra and Sanjaya are engaged through partner-scoped design partnerships, with detail shared under those agreements.
Frequently asked questions
How can AI be used in defence without autonomous weapons?
The largest near-term uses are sustainment and support: maintenance and fault-finding, logistics, manuals and SOP answers, report summarisation with redaction, imagery analysis support, cyber alert triage, secure software engineering, procurement files, training content and disaster-relief coordination, each with a human deciding.
Can AI agents run in an air-gapped defence network?
Yes. The AgentAnywhere platform deploys air-gapped with AgentAnywhere Swaraj on customer-owned hardware, with no dependency on an external service, and records every call in a signed log that can be verified offline.
Does AgentAnywhere support targeting or weapons systems?
No. AgentAnywhere's agents and models are not offered for targeting, fire-control or any weapons chain, and are never placed in a safety path or control loop. They are limited to observe, inform, sustain and secure functions with a human deciding.
How is sensitive information protected when defence reports are summarised by AI?
Personnel, unit and location identifiers can be irreversibly redacted by AgentAnywhere Veil before any model reads the document, and the receipt records that a redaction happened and under which policy, without storing the redacted value.
Written by
AgentAnywhere Research
The team that builds the platform and the models
AgentAnywhere Research writes about the platform, the model families and the trust layer we build and run in India. Where a figure is ours, it says what it covers; where something is a demonstration or in preview, it says so.
