Class of Research
Produce Real
AI Research.
The most rigorous undergraduate AI research program in Pakistan. Selected researchers work under international mentors to produce original work. All research is published on arXiv, Papers with Code, and Hugging Face, with exceptional papers submitted to international venues like NeurIPS, ICLR, and ACL workshops. 20–30 seats per cohort. Highly competitive.
The work
What You'll Do
Original Research
Identify a real problem in Pakistani context, formulate a research question, and conduct original AI research under mentor guidance. Your work addresses gaps no one else is filling.
Mentor Sessions
Weekly 1-on-1 sessions with international AI researchers. Get direct feedback on methodology, experimental design, results interpretation, and paper writing.
Conference Submission
Your final output is a research paper published on arXiv, Papers With Code, and Hugging Face — real artifacts, not class projects. Exceptional papers are submitted to NeurIPS, ICLR, or ACL workshops.
How you get in
Your Selection Process
Four stages. Each one designed so that even if you aren't selected, you walk away with something you built or learned.
Application
Free · 10 minutes
Tell us why you want to do research and why now. We don’t care about your GPA or prior publications. We care about intellectual curiosity and the ability to ask good questions.
Screening
Free · 30 minutes
Two written questions testing critical thinking about AI in Pakistani contexts. AI tools are allowed — we’re testing genuine reasoning, not memorization. Can you see what others miss?
Assessment
Rs. 1,000 · 48 hours
You receive a real Pakistani problem and write a 2-page research proposal. The quality of your research question is weighted at 40%. We’re looking for researchers who can frame problems, not just solve them.
The output
What a Class of Research
output looks like
Bridging the Urdu-English Gap: Low-Resource Cross-Lingual Transfer for Sentiment Analysis in Pakistani Social Media
A. Khan*, S. Ahmed*, M. Rashid — Afrium Research, 2026
Abstract
We present a cross-lingual transfer framework for low-resource Urdu sentiment analysis, leveraging code-switched Pakistani social media data. Our approach fine-tunes multilingual transformers on a novel Urdu-English corpus of 47K annotated tweets, achieving state-of-the-art F1 scores on Roman Urdu benchmarks while requiring 73% less labelled data than monolingual baselines. We release all data and models publicly.
Ready to produce real research?
20–30 seats. International mentors. Conference submissions. Pakistan's most serious undergraduate AI research program.