Starting over may be simpler.
We help organizations start again with AI—without losing what already works. Our own research and applications show us how.
Tokenizer
Watch one sentence become a compact, legible token trace across Romanized Nepali, Devanagari, and mixed context.
Token trace
Meaning preserved across scripts.
More language. Less waste.
Nepali, Romanized Nepali, and English—held in one vocabulary without flattening the language.
A compact Nepali model,grown from scratch.
A beginning, still becoming. Built to study what a smaller language model can preserve when every decision remains visible.Read the model noteKeep the evidence.
Clear the assumptions.
Each experiment sharpens the next.
Remove assumptionsKeep evidence
Work that
compounds.
Our first project is a Nepali language foundation: tokenizer, model, and evaluation. As we add projects across industries, each one becomes proof, infrastructure, and a sharper way to build the next.
Where we are nowGrounding the system
Language foundation
Tokenizer, evaluation corpus, and Muna—the first 161M-parameter model.
Ground the system.
Verified answers, retrieval, and evaluation that makes failure visible.
Build one useful application.
Turn the language foundation into one focused experience people can actually use.
Choose the next problem.
Let evidence from this program determine what deserves to follow.
Notes.
Published when the evidence becomes more useful than the claim.
