I build systems that hear Nepali. read documents. see pictures. guide careers.
AI Engineer at System Shared Inc. in Japan, and AI team lead at Kathmandu University’s Active Learning Lab. Research on one side, AWS deploys on the other.
Computer engineer with a research background in machine learning, now shipping the production systems that research turns into.
I like taking a model from a research idea to something that actually runs in front of people — training it, wiring it into a backend, and deploying it. That has meant Nepali speech recognition, LLM document QA, image tagging, and the AWS plumbing that keeps them all alive.
The same week might involve fine-tuning a transformer and fixing the CI/CD pipeline that serves it. I am happiest when both are on the table.
System Shared Inc., Japan
Active Learning Lab, KU
Research prototypes and shipped systems, from speech models to production AI backends.
A career support platform with an interactive roadmap display and AI-powered conversational counselling. OpenAI API on the front of it, an AWS backend behind it, and the CI/CD to keep shipping.
Large-vocabulary continuous speech recognition for Nepali, built on a CNN + Transformer architecture. The work became a paper at LDK 2023 — a language with very little training data got a model that hears it.
A chatbot on Llama 2 and Vicuna that answers questions from Dutch-language documents and talks to a SQLite database in plain language.
Learns image tags autonomously from internet data and builds a tag hierarchy through word embeddings. The project I currently lead at the Active Learning Lab.
A pretrained CNN encoder with a transformer decoder, writing captions in Nepali.
Site search that suggests as you type, using nothing but the site’s own content as its index.
An online platform for listing and booking homestays, hotels and rental houses.
An Android app that tracks a phone and syncs its messages and call logs to the web in real time. Won Best Software Project at the LITE technical exhibition.
There is usually something half-built on a branch somewhere. If you want to know what I am poking at this month, just ask.
Ask meShishir Paudel, Bal Krishna Bal, Dhiraj Shrestha. Proceedings of the 4th Conference on Language, Data and Knowledge (LDK 2023), NOVA CLUNL, Portugal, pp. 328–333.
LITE Technical Exhibition, 2017
3-day NVIDIA workshop, 2021
Kathmandu University delegate — a week in Japan
There is a whole other half — guitars, ink drawings and far too much basketball. Flip the switch.
The parts of me that never make it onto a CV.
Badly on a good day, decently on a great one. Go on — pluck a string.
Pencil, ink, and a stubborn habit of redrawing the same face until the eyes finally sit right.
Basketball first, then football, then cricket — whatever gets a group of people outside.
From Pokhara. I build AI with a team in Japan, and still lead research with a lab back in Kathmandu.
Most of my research exists because Nepali does not have the datasets English does. That gap is the whole reason the work is fun.
Papers, production systems and six years of experience are one click away.
Open to conversations about AI engineering, applied research, low-resource languages, and interesting backend problems.
