Intelligence rooted in ancient wisdom
We are building language models grounded in Hindu scriptures, ancient manuscripts, and indigenous sciences — knowledge passed down for thousands of years, now precisely understood by machines.
What's new
Nithyananda AI-1: a foundation model trained on 2.4M Sanskrit verses
Apr 8, 2026
Reading the unread: OCR for palm-leaf manuscripts in Grantha and Devanāgarī
Apr 8, 2026
Why Jyotiṣa needs its own evaluation suite — proposing an open benchmark
Mar 28, 2026
Open-sourcing our Devanāgarī tokenizer: 50% smaller, 2× faster than BPE baselines
Mar 11, 2026
Our research, alive in Ask Nithyananda.
Every model, dataset, and benchmark we build flows directly into Ask Nithyananda — a conversational interface for the depth of Hindu knowledge. Verses are quoted with exact citations, śāstra is cross-referenced, and reasoning is auditable.
- Cited answers tied to original verses and chapter numbers
- Multi-script search across Sanskrit, Tamil, Hindi, Telugu
- Lineage-aware reasoning across commentarial traditions
Six domains, one continuum.
Hindu knowledge is not a single corpus. It is a network — śāstra referencing śāstra, lineage commenting on lineage. Our research respects that structure.
The Vedas, Upaniṣads, Itihāsas, Purāṇas, and Āgamas — modeled with full awareness of their commentarial layers (Śaṅkara, Rāmānuja, Madhva and beyond) so meaning is preserved across schools of thought.
A general intelligence that remembers where it came from.
The frontier of AI is being shaped by a narrow slice of the world's knowledge. We are building the part that's missing — a civilisation's worth of insight, encoded with the rigour it deserves, available to anyone who asks.
Datasets and evaluation suites released under permissive licenses, so the wider research community can build on them.
Every claim made by our models traces back to a specific verse, śloka, or commentarial reference — verifiable, not vibes.
We do not flatten Advaita into Viśiṣṭādvaita. Each darśana is preserved with its distinct epistemology.
Hindu Knowledge Eval (HKE-v1)
Average accuracy on 12,400 expert-graded questions across four domains. Higher is better.