SAMVIT Platform

Overview

India's sovereign AI platform for specialized foundational models that are frugal, secure, and deployable across cloud, enterprise, edge, and future mobile environments.

Details

SAMVIT is IAIRO’s sovereign AI platform for small AI, as opposed to big AI represented by large language models. It sits atop AI infrastructure such as IndiaAI Infrastructure and compact, desktop-class AI supercomputers, enabling frugal development and operation of specialized foundational models.

These models are task, domain, or industry specific and can be deployed across cloud, enterprise, and edge environments, with future support for mobile settings. They are designed to meet stringent security and privacy requirements.

The platform is grounded in knowledge-graph-based neurosymbolic AI pioneered by IAIRO researchers. By combining data, knowledge, and human expertise, SAMVIT supports higher-quality models with user-level explainability, safety, grounding, and reduced hallucinations for mission-critical and high-value applications.

Big AI vs. Small AI

AI has recently been dominated by LLMs, and for a while the mantra was “bigger is better.” A few companies with massive resources have cornered this market — this is Big AI. For consumer-centric tasks such as question-answering, content creation, and coding, Big AI has driven exceptional progress. But for many high-value and mission-critical enterprise applications, LLMs are not the right solution. For these, a Small AI approach is better suited — and India has the opportunity to take global leadership in this emerging form of AI.

What makes a Small AI model

Small AI models — not limited to popular Small Language Models — sit at the center of this ecosystem and are characterized by four traits:

• Neurosymbolic — data alone is not enough; these models combine data and knowledge (symbolic representation, such as a knowledge graph or logic, including rules, guidelines, and regulations).
• Custom — not one-size-fits-all, but highly targeted, domain-, industry-, task-, or enterprise-specific models, trained only on high-quality, often curated, data and knowledge.
• Compact — small or right-sized, edge-deployable or on-premise, low cost to create and operate.
• Composite — able to use tools, small models, and LLMs in a hybrid agentic system to solve complex application tasks, often with human engagement.

Small AI models support the development of intelligent, robust, and trustworthy AI. Some of the 15 foundational elements that enable this include reliability, consistency, alignment, causality, instructability, reasoning, planning, grounding, interpretability, explainability, and safety.

Current work

• Frugal training and inference pipelines for specialized foundational models.
• Knowledge-graph-grounded neurosymbolic reasoning for trustworthy deployment.
• Cloud, enterprise, edge, and future mobile deployment profiles with privacy and security constraints.

For further information, see the LinkedIn overview.

Return to Projects and Themes