ShepHertz
All news
AgentAnywhere launch visual for Tatva, Sanjaya and Drashta: the line 'Don't rent your intelligence. Own it.' beside a simulated CGI drone swarm above a city.

Platform

Big AI for small machines: ShepHertz unveils Tatva, Sanjaya and Drashta, India-built AI models that run offline on constrained devices, from satellites to smart rings

At Drone Expo 2026 at Yashobhoomi, New Delhi, ShepHertz unveiled three India-built AI model families for machines that work where the network does not: Tatva, Sanjaya and Drashta. They run offline, on the device, on constrained hardware, from satellites to smart rings. The model proposes, a person decides, and every step is signed.

Three families, one discipline

  • Tatva turns plain language (English, Hinglish or an Indian language) into exactly one command the machine already understands, or refuses. It runs on the device with no network, and it is trained from scratch in India. Tatva Edge is in private preview.
  • Sanjaya watches the telemetry of systems that cannot fail, from spacecraft ground segments to grids and plants, and flags anomalies for the operator to act on. It is engaged with design partners.
  • Drashta gives machines eyes. It finds aircraft and structures in satellite and aerial imagery, offline and inside the perimeter, and hands structured findings to an analyst, who decides. It is engaged with design partners.

A person in command

None of these models is the last word. Rules that sit outside the model decide what is allowed. A named person approves anything consequential. Every step leaves a signed receipt that can be checked offline. Tatva is not a targeting or engagement system and is never placed in a weapons-release chain.

The demonstrations shown at the expo, including drone mission operations, a humanoid assistant and a swarm handover, were simulated. The policy, approvals and receipts behind them are real code.

Opening the trust toolchain

Alongside the models, ShepHertz is opening its trust toolchain under Apache 2.0. Shuddhi, the data factory that cleans a training corpus and issues provenance receipts, is on GitHub today. Model Hub, Agent Universal Gateway, Annotator and Nabhika OS follow over the coming weeks, documented at docs.agentanywhere.ai.

Read more

Back to all news