LLM Behavior · Tool Use · Multilingual Adaptation
MSc Computer Science @ UBC/ advised by Prof. Laks V.S. Lakshmanan/ Vancouver, BC
السلام علیکم! I'm a Computer Science master's student at UBC, advised by Prof. Laks V.S. Lakshmanan. My research is about understanding why large language models behave the way they do and then optimizing that behavior into something better.
I want models that do more than sound capable: models that know when to reason, when to act, and when not to pretend they know.
How should an LLM agent balance three forces: what the user is asking for (and what it may still need to ask them), what the model already knows in its parameters (internal reasoning), and what it can reach for in the world through tools and function calls (external action)? The interesting question is when a model should think versus act versus pause to ask for clarification and how to make that trade-off efficient, reliable, and safe.
📚 Also co-author of two books: Mastering Digital Art with Stable Diffusion and Maximizing Productivity with ChatGPT and full publication list on my Google Scholar.
Out of 13 countries across APAC, our team took home the top award for building culturally-relevant LLMs for the 2.5 billion people the internet keeps forgetting — representing Pakistan on a global stage. It's the project that turned "make models work in Urdu" from a side quest into my research direction.
Joined Prof. Laks V.S. Lakshmanan's group to work on LLM behavior, agents, and tool use.
Token expansion, continual pre-training, and culturally-grounded instruction tuning; +14.4% on the Urdu MMLU benchmark.
The non-profit I founded to grow women's careers in STEM through mentoring, skill-building, and confidence. Pakistan went on to secure the 2nd-highest count of Linux Foundation Scholarship recipients worldwide.
800+ students · 8 team membersCommunity lead at Cohere For AI, an open-science non-profit. I host and coordinate study sessions with researchers around the world.
3,500+ members · 100+ countriesExtended the reach and mission of MITACS (a Canadian research non-profit) into Pakistan, connecting students to global research opportunities.
1,500+ membersWords, not just weights
Writing is how I test whether I really understand an idea: if I can't explain clearly, I probably haven't understood it deeply enough yet.
I'm looking for research internships in NLP and LLMs behavior, agents, tool use, multilingual. If your team cares about understanding why models do what they do, I'd love to talk.