arcelis labs

Research lab · Kathmandu, Nepal

From first question to first paper.

Arcelis Labs is a research lab in Kathmandu. Students work on research projects, and professors and researchers in Nepal or abroad advise them. Each project aims for a paper, code or data that others can use.

I'm a student I'm a professor

For students Work on a research problem with an experienced advisor.

For professors Advise students on a question you care about. Remote advising works.

Open by default Code, data and results are released openly where possible.

How it works

From a question to a published result.

  1. 1

    Proposed

    A professor, researcher or student submits a project idea.

  2. 2

    Reviewed

    We check that it is feasible, has a committed mentor and a clear output.

  3. 3

    Open

    The project is listed here and students apply to join.

  4. 4

    Active

    The team works to a plan, with regular meetings and deadlines.

  5. 5

    Published

    Results are submitted to a venue, with code and data released openly where possible.

What we commit to

  • Every applicant gets a reply within 14 days.
  • Credit is clear from day one. Authorship follows contribution, and we agree on it before work starts.
  • Open by default. Code, data and benchmarks are released under open licenses whenever we can.
  • Mentors from anywhere. Professors can mentor remotely, from any university in the world.

Projects

Projects open to students.

Each project has a mentor, a clear question and an expected output.

First call for projects

Our first projects are being prepared.

Professors and researchers can propose a project now. Students can register their interest and hear first when projects open.

Propose a project Register interest

Research focus

Research areas.

Our first projects are in language and computing. Proposals in other areas are welcome.

Low-resource NLP

Methods for languages with little digital text, starting with Nepali, Maithili, Tamang and Nepal Bhasa.

Benchmarks & evaluation

Open datasets and reproducible evaluations that show what models can and cannot do in these languages.

Efficient systems

Inference, kernels and HPC that let models run fast on modest hardware, offline and at the edge.

Why it matters

हाम्रो भाषा , हाम्रो अनु सन्धान ।

“Our language, our research.”

Nepali
82 bytes
English
27 bytes
Ratio
3.0×
The same sentence takes about three times as many bytes in Nepali, because every Devanagari character, including vowel signs, uses three bytes in UTF-8. Gaps like this shape how well, and how cheaply, models handle the language.

People

People.

Loading people…

Publications

Selected prior work.

  1. 2026Under review

    When Do Graphs Help Transformers? Graph Augmentation in Nepali News Classification

    Swastik Bhandari, Suman Shrestha

    A parameter-efficient BERT-GCN model for 122K Nepali news articles, reaching 92.60% accuracy.

  2. 2026Under review

    Deep Learning for News Recommendation: A Global Review with Low-Resource NLP Perspectives

    Swastik Bhandari, Suman Shrestha

    A PRISMA systematic review of 2,954 publications.

  3. 2026Conference talk

    Offline-First AI Decision Support for Maternal and Child Health

    Swastik Bhandari

    4th National Telemedicine & 1st Digital Health Conference (TMSoN), Kathmandu.

Join

Work with us.

Students

Any university in Nepal, any year. Send a short note on what you have worked on, and we will contact you when a project fits.

Register interest

Or write to join@arcelis.org

Professors & researchers

In Nepal or abroad. If you have a research question and time to advise, propose a project. We review it, list it, and help you find students.

Propose a project

Or write to research@arcelis.org