ozanbayiz?
I'm a 5th-year MS student in EECS at UC Berkeley, advised by Professor Narges Norouzi. I work on video understanding in service of AI for education.
My research interests include representation learning, multimodal integration, and data-efficient RL.
Previously, I was head TA for EECS 189/289A (Intro. to ML) and a student researcher on Google's XR Perception team.
I remain eternally grateful for Vongani Maluleke's kind mentorship and guidance, and for the support of my family and friends.
research
When Offline RL Cannot Evaluate Teaching: A Diagnostic Case Study
Decision Transformer tutoring policies over a knowledge-tracing world model, screened by a four-diagnostic OPE protocol — every comparison between pedagogically-grounded reward shapes and the immediate-correctness baseline was caught, each catch tracing to a concrete data-infrastructure gap.
Frozen Vision Encoders for Deep RL in VizDoom
Double DQN over frozen self-supervised vision features (AIMv2, V-JEPA 2) with PCA whitening — decoupling representation from policy learning makes training fast, memory-efficient, and stable.