who am i
I grew up in a farming household in rural Punjab, Pakistan. There was no obvious path from there to AI research. I found one anyway.
A scholarship brought me to Budapest to study computer science at ELTE. That was the first door. The rest I had to open myself.
I cold-emailed my way into ETH Zürich’s Agentic Systems Lab as the only bachelor’s student in a master’s-level cohort. I joined Infineon to work on multilingual LLM pipelines. I also researched bias across synthetic personas, and that work ended up at international conferences.
Somewhere along the way I noticed a pattern in my own work. I kept caring less about making AI systems impressive and more about what happened after the demo. Agents that loop and never recover. Retrieval that returns the wrong document without anyone noticing. Memory that slowly changes how a system behaves. Models that answer confidently and wrongly.
So that became the thesis: build agents, then figure out where they break.
That’s what my projects are about. Agent Autopsy analyses agent traces to explain why a run failed. My research looks at how retrieval systems degrade and how to evaluate failure instead of just detecting it. My thesis tackled grounding and hallucination in document QA.
I also like building communities around this. I co-founded ELTE’s Data Science Club and co-organised an agentic AI hackathon in Budapest with around 150 participants.
The fastest way to learn where systems break is to watch a lot of people build them under pressure.
I’m using this blog as a public notebook. I’ll write about agent failures, retrieval, memory, evaluation, and the unglamorous engineering that keeps these systems working.
If any of that overlaps with what you’re working on, reach out. I answer cold emails. I owe my career to one.