Tosh AI प्रज्ञा
Sovereign AI, retrieval and agentic orchestration platform
“Defence-grade reasoning that runs inside your own walls.”
THE PROBLEM
Defence formations, national-security agencies, ministries and regulated operators want the reasoning capability of modern AI assistants, but most of those assistants are externally hosted services. Prompts and outputs are processed on infrastructure the customer does not control, under a jurisdiction the customer may not choose. They are not designed to run inside a classified network, a forward location or a sealed government facility. And the customer usually cannot control retention, inspect what was retrieved, or prove what the system did and who approved it. The model is rarely the hard part any more. The hard part is a controlled system around the model that keeps data, access, approvals and audit under the operator's authority.
THE SOLUTION
What Tosh AI Does
Tosh AI is a sovereign AI, retrieval and agentic orchestration platform for defence, national security, government and critical systems. It brings reasoning, grounded retrieval and human-authorised action inside infrastructure the customer controls: on-premises, in an approved sovereign cloud, or air-gapped. It is not a chatbot and not a clone of any foreign assistant. Models are replaceable components. The value is the controlled reasoning and action system around them.
The platform has three parts. PRAGYA is a model-family and adaptation strategy: it uses appropriate open-weight foundations, with continued pretraining, fine-tuning, retrieval and task-specific models selected per security boundary, language, hardware and task. It does not depend on training a frontier model, and a better foundation can be adopted without surrendering the customer boundary. Sovereign retrieval grounds answers in the customer's own documents, inside their own perimeter, with citations. YANTRA, the orchestration layer, breaks a goal into visible steps, calls only allow-listed tools, limits access by identity, role, data boundary and task, and applies an approval policy before any consequential action. Every plan, retrieval, tool call, model output, approval and action is attributable.
The public assistant at tosh.ai is a public trust product that demonstrates usability and gathers feedback. It is separate from customer-controlled deployments and is not evidence of air-gapped operation. Model sizes, supported hardware, latency and task performance are confirmed per deployment before any proposal.
प्रज्ञा (Pragya) - wisdom, discernment, applied intelligence. PRAGYA is the model-family and adaptation strategy inside Tosh AI, built on replaceable foundations. YANTRA is the orchestration layer that turns reasoning into steps a human can see and approve. PRAGYA reasons. YANTRA orchestrates under human authority.
CAPABILITIES
Key Features
PRAGYA Model Family
A model-family and adaptation strategy on replaceable foundations, often open-weight, adapted through continued pretraining, fine-tuning and task-specific models. The intended tiers are Edge, Core and Max. The base model can change as the open ecosystem improves, without changing where your data lives.
Sovereign Retrieval with Citations
Answers are grounded in your own documents, inside your own perimeter, with citations back to the source. Retrieval respects the access rights of the person asking. For defence and government work, an ungrounded answer is a liability; a cited, reviewable one can be checked.
YANTRA Orchestration
Plans are decomposed into visible steps. Tools are explicitly allow-listed. Access is limited by identity, role, data boundary and task. YANTRA coordinates retrieval, reasoning and tool use for multi-step work, and fails safely when data or confidence is insufficient.
Human-Authorised Action and Audit
Consequential actions follow an approval policy set by the customer: the system proposes, an authorised person decides. Every plan, retrieval, tool call, model output, approval and action is logged and attributable, and answers expose their sources, uncertainty and limitations.
Customer-Controlled Deployment
Model, retrieval and orchestration run inside your perimeter, on-premises, in an approved sovereign cloud, or air-gapped with no mandatory network egress. Containerised for Docker and Kubernetes. Hardware and deployment scope are confirmed per engagement.
Language and Domain Adaptation
Adaptation to the customer's languages, including Indic languages, and to their doctrine, technical manuals and working vocabulary, so the system is useful for the work that matters rather than for generic benchmarks.
Edge and Disconnected Operation
PRAGYA Edge is the intended tier for small, quantised models in disconnected settings such as a forward location, a vehicle or an isolated workstation. Disconnection is treated as a normal operating condition, and offline behaviour is defined per deployment.
No Mandatory External Runtime Dependence
No external analytics, no externally hosted models and no tracking SDKs in customer deployments. Data is stored and processed inside the customer's jurisdiction or perimeter, with documented interfaces and export paths so the customer can inspect, migrate and continue operating.
DEPLOYMENT OPTIONS
Product Tiers
PRAGYA Edge
Intended for small, quantised models in disconnected and forward-deployed use, where connectivity cannot be assumed. Tier names describe architecture and roadmap; supported hardware and model sizes are confirmed per deployment.
PRAGYA Core
Intended for on-premises operational workloads: grounded retrieval, reasoning and everyday YANTRA workflows inside your perimeter. Capability on your hardware and tasks is confirmed before a proposal.
PRAGYA Max
Intended for larger customer-controlled clusters or an approved sovereign cloud, for complex reasoning and larger-context analysis. Model selection, latency and evaluation evidence are confirmed per deployment.
CUSTOMER CONTROL COMPARED
Why Tosh AI?
“Tosh AI is not a chatbot. Models will keep improving and keep changing. What a defence or government operator needs is the system around them: grounded retrieval, visible steps, allow-listed tools, human approval and an audit trail they own. This is defence-grade reasoning that runs inside your own walls.”
INTEGRATION
Works With
DESIGNED FOR
Who It Serves
Built in India. Designed for Sovereign Nations.
Engineered in India for lawful deployment by sovereign nations and approved organisations. Customer-controlled infrastructure, documented interfaces and export paths, and no mandatory runtime dependence on infrastructure outside your boundary.
EXPLORE MORE
Other Mission Applications
SECURE INFORMATION EXCHANGEसंकेतAvailableSANKET
Sovereign Authenticated Network Key-based Encrypted Transmission
Sovereign secure information exchange platform for military units and high-trust organisations, deployed and governed inside the customer's own boundary.
Learn More
RECORDS & COMPLIANCE INTELLIGENCEअभिलेखCommercially operatingRecordsKeeper.AI
Sovereign AI-First Records & Data Management with Intelligence Platform
AI-first records, knowledge and compliance intelligence that turns organisational records into a searchable, governed and auditable knowledge system.
Learn More
ASSET READINESSसजगIn active developmentSAJAG
Smart Asset Judgement & Availability Guard
Predictive maintenance and spares-demand intelligence that estimates failure risk from historical maintenance, usage, terrain and weather data.
Learn More
INFRASTRUCTURE AWARENESSदृश्यIn active developmentDRISHYA
Defence-grade Real-time Integrated Single pane for Holistic Yield Analytics
Multi-OEM infrastructure monitoring that brings telemetry and alerts from heterogeneous network equipment into one operational picture.
Learn More
CYBER RESILIENCEमायाIn active developmentMAYA
Military Autonomous Yield-intelligence Architecture
Defensive cyber deception and incident intelligence: approved honeypots, honeytokens and canary artefacts that expose intrusions and give operators context.
Learn MoreReady to See Tosh AI in Action?
Start with a controlled demonstration scoped to your operational problem, security boundary and existing systems. Available to sovereign and approved organisations, subject to applicable export law.