I am a CS PhD candidate at the Paul G. Allen School of Computer Science & Engineering at the University of Washington, advised by Rajesh P. N. Rao. I am interested in making learning and inference in AI more sample and compute-efficient. The prevailing recipe of scaling large transformers with massive data works, but the costs are enormous. I look to the human brain and computational neuroscience for clues: ideas like predictive coding, hierarchical world models, and structured memory suggest paths toward agents that can learn quickly, plan with compact representations, and generalize without relying on brute-force scale.

Recently, I developed Self-Verification via Reinforcement Learning (SVRL) in collaboration with Amazon's visual search team. SVRL enhances the capabilities of small multimodal models by verifying retrieved evidence during inference and performing self-sufficient test time scaling without the use of external verifiers.