ShepHertz is developing India-built AI models intended to run offline on constrained devices. Tatva, Sanjaya and Drashta target commands, telemetry and visual analysis, while the company’s emphasis on human approval and traceable records aims to address trust. Independent performance testing and broader availability will determine their practical impact.
Indian technology company ShepHertz has introduced three AI model families designed to run on devices with computing resources, including systems that cannot rely on a constant internet connection. This announcement addresses a key AI challenge: making it useful when cloud access is unavailable, expensive, or unsuitable.
Make Telecom Talk My Trusted Source
Key Highlights
ShepHertz introduced Tatva, Sanjaya and Drashta for on-device AI use cases.
The models target command processing, telemetry monitoring and machine vision.
The company says it is opening its AI trust toolchain, beginning with Shuddhi.
Shephertz Wants Offline Native AI to Power Devices
Much of today’s AI experience depends on sending requests to servers over the internet.
That approach works well in many environments, but it can limit equipment operating in remote locations, industrial sites, or other places where connectivity is unreliable.
ShepHertz says its new model families, Tatva Sanjaya and Drashta, are designed to bring AI to the machines they run on.
The company unveiled them at Drone Expo 2026 in New Delhi as part of its AgentAnywhere platform. The company says it built the models in India and can run them offline on constrained hardware, with applications ranging from satellites and drones to smart glasses and smart rings.
Shephertz New AI Models Explained
Each model family has a purpose. Tatva translates plain-language instructions into commands a machine already recognizes.
The company says it can handle English, Hinglish, and Indian languages, but it refuses instructions outside the permitted command set. Tatva Edge is currently in preview, so its broader availability and real-world performance remain to be established.
It is intended to flag patterns for operators to review, with potential uses in infrastructure, industrial systems, and other settings where early warning is critical. ShepHertz says Sanjaya is working with design partners.
Drashta is the company’s machine-vision model family. It is designed to analyze satellite and aerial imagery, identify objects such as aircraft and structures, and provide findings for human review.
The company also lists this model as engaged with design partners rather than generally available.
How Does Offline AI Approach Help?
The offline approach could be useful in settings where data cannot easily leave a site or where network connections are intermittent. For example, an industrial operator may want equipment monitoring to continue during a connectivity outage, while a wearable device may benefit from local processing.
These are applications, however, not proof that every model is already deployed in consumer products such as smart rings or glasses.
ShepHertz is also emphasizing supervision and traceability. The company says its systems are designed so a model proposes an action, a person approves the decision, and each step generates a signed catalog that can be checked offline.
This distinction counts when AI is used around infrastructure or machinery: detecting a pattern or suggesting a command should not automatically authorize the system to operate.
Alongside the model announcement, ShepHertz says it is opening parts of its AI trust toolchain under the Apache 2.0 license.
The first component, Shuddhi, is described as a data-processing tool that cleans training datasets and creates provenance records. The company says it plans additional components, but their availability should be checked against its updates.
The announcement is notable because it looks beyond chatbots and cloud-based assistants. Smaller specialized models could make AI more practical for devices that have limited power, memory, or connectivity.
The real test will be technical evidence of the models’ accuracy, the hardware they require, their response speed, and how well they handle errors or unexpected instructions.
Challenges on Running Offline Based AI System in India
For India’s AI ecosystem, the effort also raises a question about where intelligence runs and who controls the data. Models built to operate within a customer’s own systems may offer more control in some use cases.
That alone does not guarantee security or reliability; organizations should test them independently before relying on them for decision-making.
For now, ShepHertz has outlined a set of models aimed at making offline AI possible across different types of equipment. The next milestone is to demonstrate how well they work outside of launch presentations and controlled trials.
Some people read for free. A few choose to support. If you found TelecomTalk useful, you can help keep us running.
FAQs
What is ShepHertz developing?
ShepHertz has introduced three specialised AI model families under its AgentAnywhere platform, designed to perform tasks on devices and systems with limited computing resources.
Which AI models has ShepHertz announced?
Tatva handles predefined language-to-command tasks, Sanjaya focuses on telemetry and anomaly detection, and Drashta analyses satellite and aerial imagery.
Can these AI models work without the internet?
The company says the models are designed for offline operation on compatible hardware. Their actual capabilities depend on the model, device and implementation.
Will ShepHertz’s AI models come to smart glasses and rings?
The models are designed for constrained devices, including smart glasses and smart rings. However, widespread consumer availability depends on manufacturers integrating and deploying them.
How is offline AI different from cloud AI?
Offline AI processes supported tasks locally on a device or compatible system. Cloud AI sends requests to remote servers for processing. Local processing can reduce connectivity requirements, while cloud systems can access more powerful computing resources.
Support Telecomtalk
