India AI Impact Summit Insights
All sessions

Gyan Bharatam: Where Tradition Meets Technology

Contents

Executive Summary

The Gyan Bharatam initiative represents a landmark effort to digitize, preserve, and democratize India's vast manuscript heritage using AI and machine learning technologies. Rather than replacing tradition, the project harnesses cutting-edge AI—including OCR, handwritten text recognition, large language models, and multilingual translation systems—to unlock centuries of civilizational knowledge embedded in Sanskrit, regional scripts, and ancient texts that remain scattered, fragile, and largely inaccessible. This collaborative mission aims to transform India's knowledge legacy from museum artifacts into living resources that drive innovation, scholarship, and cultural continuity into 2047 and beyond.

Key Takeaways

  1. Gyan Bharatam is not digitization alone—it is a knowledge extraction, authentication, and democratization ecosystem combining OCR, LLMs/SLMs, speech technology, and semantic search to make ancient wisdom actionable for modern India.

  2. Indian AI must be built on Indian knowledge, not imported frameworks. Manuscript corpora provide finite, authentic data that create genuine competitive advantage in building culturally-grounded, ethically-sound AI systems.

  3. Preservation does not mean curation or control: The federated model protects original manuscripts while extracting and openly sharing knowledge through multiple languages, scripts, and modalities—balancing conservation with accessibility.

  4. Solving the script problem is solving the knowledge problem: Investment in region-specific, temporally-aware, context-preserving OCR and NLP is foundational. Generic Western tools will fail; custom solutions are not a limitation but a necessity and opportunity.

  5. This is a circular loop: Manuscripts inspire AI; AI-enabled AI systems should be trained on manuscript-derived values (Dharmic AI) to avoid perpetuating Western cultural biases in the systems India builds for its citizens.

Key Topics Covered

  • Manuscript Preservation & Digitization: Physical conservation, image enhancement, AI-enabled restoration, and creation of machine-readable formats
  • Script & Language Complexity: Regional variations in historical scripts (Devanagari, Brahmi, etc.), script mixing, language identification challenges, and temporal evolution of writing systems
  • Optical Character Recognition (OCR) & Handwriting Recognition: Limitations of adapting Western OCR engines to Indian languages; need for script-specific and region-specific algorithms
  • Knowledge Extraction & Authentication: Challenges in maintaining contextual integrity, epistemological frameworks, and avoiding decontextualization of Sanskrit and classical texts
  • Large Language Models (LLMs) & Small Language Models (SLMs): Development of domain-specific, manuscript-rooted AI models (Bharat GPT, Ayurveda SLM, Ganit SLM, Gandhastra SLM)
  • Multilingual & Multimodal Access: Voice-based and text-based querying, automatic speech recognition (ASR), and translation across 22+ Indian and 35+ international languages
  • Scholarly Tools & Infrastructure: Semantic search, metadata extraction, concept-based retrieval, cross-manuscript searching by subject domain
  • Stakeholder Collaboration: Partnerships between academia (IIT, Sanskrit University), government, technology companies (Inverse AI, Bhashini), cultural institutions, and international repositories
  • Cultural & Epistemological Framework: Importance of oral tradition, contextual reading, lived practices, and Dharmic AI vs. Western-centric AI ethics
  • Scientific Knowledge Recovery: Hidden scientific knowledge in manuscripts (chemistry, medicine, astronomy, mathematics, metallurgy) that originated in India but was later attributed to Western sources

Key Points & Insights

  1. The Accessibility Crisis: Thousands of years of Indian scientific, philosophical, and practical knowledge remain trapped in fragile manuscripts written in scripts few people can read today—a loss that limits both scholarly understanding and contemporary innovation.

  2. Script Complexity Cannot Be Outsourced: Unlike English OCR, Indian manuscript recognition requires understanding regional variations, temporal evolution of letterforms, script mixing, language ambiguity, and material conditions. Western algorithms fail; custom, India-specific solutions are mandatory.

  3. Authentication Through Lineage: Sanskrit scholarship traditions validate knowledge by consulting original manuscripts when sources contradict—a practice now being digitally systematized. AI must operate within this epistemological framework, not replace it with Western-style confidence scoring.

  4. Context is Everything: A single Sanskrit word decontextualized loses meaning. Machine understanding of Indian texts requires encoding oral tradition (pronunciation and recitation practices), textual context, and lived cultural practices—not just word embeddings.

  5. Circular Empowerment: AI Needs Tradition, Tradition Needs AI: AI can extract and translate knowledge from manuscripts AND derive ethical, cultural values from that knowledge to build "Dharmic AI" that reflects Indian civilization's values—breaking dependency on Western-centric AI systems.

  6. Knowledge Sovereignty & Data Reclamation: India generates 15-18% of world's data yet imports tokens from Western LLMs. Manuscript-grounded LLMs create genuine competitive advantage: finite, authentic, sourced knowledge that cannot be scraped from the public internet.

  7. Scattered Custodianship Requires Federated Architecture: Manuscripts are held by temple trusts, universities, state repositories, international collections (Pennsylvania, Indonesia, Iran). Preservation of originals is non-negotiable; the Delhi Declaration commits custodians to sharing knowledge, not physical items or even images.

  8. Domain-Specific SLMs Over Generic LLMs: Six small language models already operational (Ayurveda, Ganit/Mathematics, Gandhastra/Perfumery, etc.). SLMs provide manuscript-referenced answers with practical utility for daily use cases, unlike generic models prone to hallucination.

  9. Democratization Through Multimodal Access: Voice-based interfaces in tribal languages (Bili, Miso, Kashmiri, Bodhi/Ladakhi) without scripts; text-to-speech; concept-based search—making knowledge accessible across literacy levels and regional/linguistic boundaries.

  10. Knowledge Reclamation Corrects Historical Misattribution: Panini's grammar → Chomsky's generative grammar → IBM AI; atomic theory from Maharishi Kanada → modern physics. Properly sourced manuscripts restore Indian origins and inspire indigenous innovation trajectories.

Notable Quotes or Statements

"AI is not here to replace tradition. AI is here to unlock it. When machine intelligence meets civilizational intelligence, something transformative happens."
— Opening remarks, Gyan Bharatam mission vision statement

"We move from preservation to participation, from archives to analytics, from storage to scholarship and from heritage to human capital."
— Secretary, on the mission's intent

"The character of each language, the script of each language has similarities as well as differences. You cannot simply take an engine used for English and adapt it for Indian languages—it's not possible."
— Professor Sinha (paraphrased), on OCR limitations

"When the knowledge is digested by the machine, how do you authenticate it? This is the tradition of India—we consult the oldest manuscript."
— Vice Chancellor, Central Sanskrit University, on validating AI-generated knowledge

"Unless we solve this problem, we will not be able to convince our own people. People living in Bharat need to see that the knowledge is rooted in a manuscript."
— Ramakrishna, Inverse AI, on why manuscript-referenced LLMs matter

"Virasat and Vikas are not decoupled. Heritage and progress are tied together."
— Rishi, Bharat Gen, on why building Indian AI requires Indian knowledge sources

"Data is finite. All LLMs have scraped the internet. Gyan Bharatam gives us genuine, sourced data no one else has."
— Ramakrishna, on competitive advantage of manuscript corpora

"We are not giving the original manuscript to anyone. The knowledge will be shared, but the manuscript rests with its custodian."
— Vice Chancellor, on federated preservation model

Speakers & Organizations Mentioned

Government & Academic Institutions:

  • Secretary, Ministry of Culture (chair and primary moderator)
  • Vice Chancellor, Central Sanskrit University (VC)
  • Professor Sinha (OCR and Indian language NLP expertise)
  • Professor Ja (Sanskrit manuscripts and science history)
  • IIT Hyderabad (represented by Professor Ravi, manuscript reading research)
  • IIT Mandi (represented by Professor Olit Saluja, consortium member)

Technology Partners & AI Organizations:

  • Inverse AI: Stitch together end-to-end solution; OCR, search, LLM integration
  • Bharat Gen: Nonprofit, government-funded foundational AI company; building culturally-grounded AI systems; multiple IIT partnerships
  • Bhashini: Multilingual translation, automatic speech recognition (ASR), text-to-speech; supports 22 Indian languages on voice, 36 on text, 35+ international languages
  • Bharat GPT (Inverse AI/Bharat Gen initiative): Domain-specific LLM rooted in Sanskrit manuscripts with manuscript references
  • Utcharanika tool (Bharat Gen/Rishi): Sanskrit pronunciation and speech learning interface

Cultural & Knowledge Custodians:

  • Temple trusts
  • State government repositories
  • International repositories (University of Pennsylvania Van Pelt Library, Indonesian archives, Iranian embassy collections)

Participants/Contributors:

  • Rishi (Bharat Gen, tool developer)
  • Ramakrishna (Inverse AI, Bharat GPT lead)
  • Amitab (Bhashini integration lead)
  • Kamill (pavilion/events coordinator)
  • Anirudan (Gyan Bharatam project director)

Technical Concepts & Resources

AI/ML Techniques & Models:

  • Optical Character Recognition (OCR): Handwritten text recognition; script-specific algorithms required for Devanagari, Brahmi, regional variants
  • Large Language Models (LLMs): Generic models; prone to hallucination without grounding
  • Small Language Models (SLMs): Domain-specific, fine-tuned models with manuscript references (Bharat GPT architecture)
  • Automatic Speech Recognition (ASR): 22 Indian languages + 35 international languages supported
  • Text-to-Speech (TTS): Multilingual synthesis with phonetic accuracy for Sanskrit
  • Semantic Search: Concept-based retrieval across documents (e.g., searching "agriculture" retrieves documents on farming, fields, crops)
  • Metadata Extraction: Colophone data (manuscript creation date, author, scribe, origin, writing conditions)
  • Image Enhancement & Digital Restoration: AI-enabled repair of damaged, faded, termite-eaten manuscripts without physical intervention

Knowledge Domains (Domain-Specific SLMs Demonstrated):

  1. Ayurveda: Charaka Samhita, Sushruta Samhita, Ashtanga Hridayam
  2. Ganit (Mathematics): Bijaganita, Lilavati, Arabia texts
  3. Gandhastra (Perfumery & Chemistry): ~10,000 formulas/recipes for creating perfume varieties
  4. Jyotisha (Astronomy & Astrology) — (mentioned, status unclear)

Languages & Scripts Supported (Current/Planned):

  • Indian languages (voice): Hindi, Sanskrit, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, Punjabi, Bengali, Assamese, Odia, Urdu, Maithili, Dogri, Konkani, Sindhi, Bodo, Mizo, Kashmiri, Bili (tribal), Bodhi (Ladakh), and others (22 total mentioned, expanding)
  • Indian languages (text): 36 languages
  • International languages: 35+ on voice and text (French mentioned as example)
  • Historical scripts: Devanagari (various periods: 6th–10th centuries AD), Brahmi, regional variants, Sharda, Granta

Infrastructure & Datasets:

  • Gyan Bharatam Pavilion: Hall 14 (later referenced as Hall 17), physical demonstration of end-to-end pipeline
  • Gan Setu Competition (September 2025): Hackathon challenging technologists to solve manuscript digitization challenges; 4 companies selected
  • Delhi Declaration (September 2025 conference): Formal commitment by custodians and repositories to collaborative knowledge sharing model

Methodological Frameworks:

  • Epistemological Grounding: Encoding oral tradition (recitation), textual context, and lived practices into AI systems
  • Federated Preservation: Originals remain with custodians; knowledge extracted and shared; images not publicly distributed without permission
  • Manuscript-Rooted Grounding: Every AI-generated answer linked to source manuscript, defeating hallucination and establishing authenticity
  • Dharmic AI Concept: Building AI systems that reflect Indian cultural values, epistemology, and ethical frameworks (alternative to "Responsible AI" or "Ethical AI" frameworks derived from Western contexts)

Related Concepts:

  • Panini Grammar System: Historically precise computational linguistics system; model for rule-based AI approaches
  • OCR Challenges: Script evolution over time (letterforms change century to century), regional variations, script mixing, language ambiguity (e.g., Kaithi could be 5-6 different languages), material degradation, termite damage

Note on Transcript Quality: The provided transcript contains significant repetition and some transcription errors (likely from automated speech-to-text processing), but the core themes, speaker identities, and technical details have been preserved and organized for clarity.