I am an M.S.E. student in Robotics at the University of Pennsylvania. Previously, I studied Computer Engineering at the University of British Columbia.
I am broadly interested in AI systems for multimodal and embodied intelligence. My work spans foundation-model infrastructure, multimodal learning, and robot learning.
Introduces a text-heavy visual instruction-tuning strategy that transfers instruction-following ability and domain knowledge across modalities, matching vision-heavy mixtures across 12 benchmarks with as little as half the training tokens.
Introduces a fine-grained concept-bottleneck framework that aligns learnable attribute tokens with localized visual evidence, enabling spatially grounded and clinically interpretable skin-lesion diagnosis.
Brings learnable community discovery to connectomic analysis by coupling transformer representations with token clustering, replacing fixed brain partitions with data-driven functional communities.
* Equal contribution
Built an AI infrastructure system for Azure OpenAI that improves inference efficiency by diagnosing prompt-cache performance at scale and translating system signals into actionable optimization strategies.
Built ClickCut, a multimodal planner-executor agent that converts creative intent into structured, executable video edits. Orchestrated multiple foundation models through tool calling and replayable workflows, making complex editing tasks reliable, inspectable, and easy to refine.
Co-founded Walnut, an AI learning platform that helps international students follow lectures through real-time transcription, translation, and context-aware study assistance. Led the technical direction across multimodal note-taking, retrieval over course materials, and the product infrastructure that turns live classroom content into searchable, personalized study workflows.
Developed a PPO-based autonomous racing policy with gate-aware observations, reward design, and domain randomization for robust high-speed flight. Bridged simulation and physical deployment on Crazyflie with ROS 2, placing 1st out of 39 teams.
Built an end-to-end edge robotics system on NVIDIA Jetson Nano, combining YOLO perception, TensorRT inference, and autonomous manipulation for real-time component sorting. Designed modular perception, control, and monitoring layers to support reliable deployment and rapid debugging.
Built a suite of robotics learning systems that connect estimation, mapping, reinforcement learning, neural rendering, and model-predictive control across complete perception-to-action pipelines. Implemented and evaluated these methods in simulation, with an emphasis on how learned policies and classical robotics components interact.