Kathmandu, Nepal — AI Fellow, Fusemachines

Unique Shrestha

I teach machines how to learn; which mostly means correcting them a thousand times until they stop making the same mistake. So, parenting, but with more math.

Unique Shrestha
हिमाल सपना शब्द चेतना … nepali morphemes Grammar-Constrained FST — ParadigmTok Character-Level Transformer Devkota v0.7 · ~15–20M params Add & Norm Generation autoregressive
Devkota — a from-scratch transformer for Nepali poetry, blocked on ParadigmTok's tokenizer until the FST ships clean.
हिमालको काखमा जन्मन्छ एक सपना,
शब्दै शब्दमा बग्छ त्यो नयाँ चेतना।
— generated output, Devkota v0.7 (draft)
01

What I actually do

I am trying to learn AI and find out: Will AI replace us? What is AI? What is it capable of? How is our data being used for training? How should it be trained? Will AI solve the problems humans create, or will AI create problems humans never intended? I am scared of AI but even more curious about it. I feel it may be the next big thing or the last big thing.

0.74
F1 - Seismosafe Nepal
100+
Stars - nepali-datasets
3
Merged PRs - scikit-learn, statsmodels, pgmpy,
02

Work

2024
Seismosafe Nepal
Building-damage severity prediction from the 2015 Gorkha earthquake dataset. THE CATCH: found and fixed a geo-target leakage bug that was inflating my own out-of-fold F1 before I trusted the number. 40+ engineered features, 8 algorithms compared, 0.74 final F1.
Feature engineeringModel selectionLeakage debugging
2025
Drug Safety Predictor
Random forest classifier estimating hospitalization risk from clinical trial data. THE CATCH: feature-importance output built in from day one, because a clinical model nobody can interrogate is a model nobody should trust.
Healthcare MLInterpretability
2025–26
Devkota
A from-scratch, character-level transformer aimed at generating Nepali poetry in the style of Mahakavi Laxmi Prasad Devkota. THE CATCH: deliberately blocked on ParadigmTok's tokenizer - I'd rather wait for a linguistically honest tokenizer than fake progress with a bad one. v0.7 stands at ~15–20M parameters.
Deep learningNepali NLPalways work in progress - by design
IOST-ASCOL
doc2tex
DOCX ↔ LaTeX converter, co-created and maintained. 56+ stars.
Thanks to @abhyudayapokhrel
Open source
IOST-ASCOL
nepali-datasets
The largest curated collection of Nepali machine learning datasets on GitHub; built because the datasets didn't exist and Nepali NLP needed them to.
Open source100+ stars
03

Research

In progress
Open source contributions
  • scikit-learn - dataset caching efficiency from 100 MB+ to 17 MBPR #33118, merged
  • statsmodels - dead code removalPR #9732, merged
  • pgmpy- DBN Inference Bug resolved PR #2702, merged
04

Writing

Loading posts…

05

Let's talk

Open to research collaboration, open source work, or a conversation about NLP, politics, philosophy,movies, or anything adjacent.

mail@uniquestha.com.np github.com/un1u3