Omeed Tehrani

I'm a researcher and engineer focused on deep reinforcement learning and large language models, with interests spanning AI safety & alignment. My current focus is on efficient ML systems, distributed training optimization, inference acceleration, and the systems-level challenges of scaling frontier models. Long-term, I want to go deeper down the abstraction stack, working on training efficiency, optimization algorithms, and the fundamental mechanisms that make intelligent systems work, whether that's at a frontier lab, research institution, or wherever the most interesting problems are.

I hold both MS and BS degrees in Computer Science from UT Austin, where I conducted research in the RobIN Laboratory(Robotic Interactive Intelligence) on transfer learning, multi-task RL using the MetaWorld manipulation benchmark suite, and return-conditioned sequence modeling with Decision Transformers on robomimic datasets, advised by Dr. Roberto Martin-Martin and visiting scholar Dr. Fernando Fernández Rebollo. I also worked in the AMRL (Autonomous Mobile Robotics Laboratory) with Dr. Joydeep Biswas on inverse kinodynamics for autonomous vehicle drifting. My work was selected for presentation at an Amazon AI Symposium.

My journey through reinforcement learning has been non-linear. Starting with robotic manipulation and policy learning in the RobIN lab, I built intuition for how agents learn from interaction, credit assignment, and the exploration-exploitation tradeoff. Those fundamentals from training robotic arms now inform how I think about LLMs as RL agents, applying insights from offline RL and sequence modeling to distributed training infrastructure, model optimization, and efficient inference at scale. I'm somewhere past the Valley of Despair on the Dunning-Kruger curve, where the derivative is finally positive again, climbing toward actual competence one paper at a time.