Gnani Artha’, a sovereign AI stack
Why in News?
The Bengaluru-based voice AI startup Gnani.ai officially launched ‘Gnani Artha’, an indigenously developed, end-to-end sovereign Artificial Intelligence (AI) stack designed for Indian enterprises and public institutions. Officially Launched by the Vice President of India, C.P. Radhakrishnan, in New Delhi on August 28, 2026.
Components of the Stack
The stack packages two primary proprietary offerings into a single integrated intelligence and execution layer:
- Evon v3.3 (The Intelligence Layer):
- An open-weights foundation Large Language Model (LLM).
- Built with 30-billion parameters using a 3-billion active parameter architecture for enhanced cost efficiency.
- Trained natively across 11 Indian languages using Indic-language datasets and domain-specific tokens.
- Gnani Plexus (The Execution Layer):
- An enterprise-grade agentic AI orchestration platform.
- Connects the underlying LLM intelligence with real-world institutional workflows.
- Automates goal-oriented tasks across specialized industries.
Key Features & Unique Advantages
- Complete Data Control:
- Runs entirely on an organisation's own infrastructure (internal data centres or Virtual Private Clouds).
- Ensures critical consumer and institutional data never leaves the local network or country.
- Elimination of the "Language Tax":
- Employs an India-focused tokenizer calibrated for regional accents and code-switching (mixing languages).
- Consumes 40% fewer tokens than traditional frontier models, massively lowering operating costs for Indian languages.
- Built-in Governance:
- Features native guardrails, observability, and audit logging within the orchestration layer rather than adding them as post-build patches.
- Allows optional human-in-the-loop validation for secure workflows.
- Sector Calibration:
- Fine-tuned specifically for highly regulated spaces including BFSI (Banking, Financial Services, and Insurance), healthcare, telecom, and government applications.
- Integrates high-performance Indic Automatic Speech Recognition (ASR) trained on real telephonic audio alongside context-aware neural speech synthesis.
