I am a machine learning researcher who keeps circling back to the same question from different
angles: not just what a model can do, but whether anyone, including the model itself, can be held
to account for it.
I completed my M.Sc. in Computer Engineering at the
University of Tehran as the
top-ranked graduate of my cohort. My thesis, supervised by
Dr. Mehdi Modarressi and
Dr. Mohammad Amin Sadeghi,
asked whether a generative model could recover what video compression throws away. The answer became
NU-Class Net, a restoration network that recovers low-bitrate video well enough to cut
file sizes roughly sevenfold without a viewer ever noticing the difference, published in
Engineering Applications of Artificial Intelligence.
That same trade-off, what to strip away and what has to survive, followed me to the
Qatar Computing Research Institute,
where I spent a year as a core contributor to Fanar,
an Arabic-centric multimodal generative AI platform. I led the effort to make its image generation
culturally legible: teaching a model to render Arabic-Islamic visual culture the way an insider would
recognize it, not the way an outsider would guess it. There was no existing yardstick for that, so I
designed fifteen LLM-as-a-judge metrics for cultural accuracy and generation quality, then checked every
one of them against Elo rankings from real human pairwise comparisons before letting the rest of the
team rely on them.
Since May 2026 I have been working with Dr. Mehrdad Tajbakhsh on a harder version of the same problem:
agents that act on their own, and the accountability trail they leave behind. The work has three
threads I actually care about: whether an agent can be made to explain and defend a decision after the
fact; what happens to that accountability when the AI infrastructure underneath sits outside the
institution trying to govern it, a question I have come to think of as Sovereign AI; and, more
concretely, building an open-weight LLM that can act as a co-expert in intangible-asset valuation,
grounded in real domain knowledge through retrieval, without leaking the private information it was
trained on.
Day to day, I split my time between two labs at the School of ECE
in Tehran. In the Computer Vision Lab, under
Dr. Mostafa Tavassolipour,
I am building a system that proposes carpet patterns for industrial weavers, not by generating an image
and hoping it survives production, but by pairing a generative model with algorithmic constraint
enforcement so every suggestion is already loom-ready: pixel-exact, on a limited palette, nothing a
machine cannot actually weave. In the
Intelligent Architectures for Computing Systems Lab,
under Dr. Mehdi Modarressi, I do the opposite kind of shrinking: pruning, quantizing, and redesigning
deep models so they fit on wearable health devices and other hardware that was never built with deep
learning in mind.
I like problems that refuse to stay theoretical. Alongside research I have shipped computer vision into
production, founded a marketplace startup, taught ten university courses, and written and filmed
explanations of the ideas I find beautiful. I was admitted twice to the Ph.D. program in Computer
Science at the University of California, Irvine (Fall 2024 and Fall 2025) but could
not enroll because of U.S. visa processing delays. So I am actively looking for a Ph.D. position where
this work, all of it, can continue.
“He who knows best, knows how little he knows.”