ciao, i am basab.

ml/ai engineer . founder . researcher .

i build ai systems that run in production, from training the models to the backends and frontends around them. i founded SAGEA, an ai research company building frontier ai from nepal. our identity-verification engine, Helios, verifies thousands of users a day at payment providers including khalti and multiple other institutions . before that i was the founding engineer at likeo (acquired for $1m) and co-founded apparatus.ai (acquired). i lead ai engineering at equalbyte, and i publish research on language models.

what i've been up to:

founder/ceo at SAGEA , an ai research lab building sovereign ai and ai infrastructure in nepal
Helios is live with khalti and multiple other finance institutions in Nepal, verifying user identity and KYC. verifies thousands of users a day
led SAGEA to join nvidia inception to accelerate sovereign ai in nepal
published the research behind it: SAGE-32B, SAGE-nano and more, all on arxiv
started stratigo with a couple of friends, to modernize microfinance institutions. here as a cto.
was a founding engineer at likeo , a virtual tryon app that got acquired by KiranaPro for $1m
before that, apparatus.ai (acquired), which was a business intelligence and forecasing startup
oh, and also, hiring at sagea.space/careers

things i've worked on:

papers:

SAGE-32B: agentic reasoning via iterative distillation (2026): a 32b-parameter model built for agentic reasoning and long-range planning, trained with iterative distillation. first author.
thinking about thinking: SAGE-nano's inverse reasoning (2025): chain-of-thought makes models reason well but leaves their decisions opaque. this makes the model explain its own reasoning. first author.
fragile mastery: are domain-specific trade-offs undermining on-device language models? (2025): what small models give up when you tune them for one domain and run them on limited hardware. first author.
the rosetta paradox (2024): models that are expert in specialized fields can fail at everyday general knowledge. first author.
the babel effect: why large language models perform differently across languages.