Building a reinforcement learning pipeline that post-trains a generative world model (NVIDIA Cosmos) to produce expert-level surgical trajectories: constructing a reward model from real expert quality scores and optimizing the policy against it via RLHF and LoRA across a multi-GPU distributed training system on PARCC HPC clusters. Generated and evaluated 120-frame rollouts with text and kinematic prompts.
Engineered a markerless 3D perception and automated data engine to extract kinematic behavior and scene understanding from monocular endoscopic video, combining vision transformers and kinematic cross-attention with classical 3D-reconstruction techniques (SfM, optical flow, state estimation). Eliminated a 31.8% frame dropout caused by occlusion and blood specularities, and cut annotation time from hours to minutes per clip, scaling the training set to 573 analyzed stitches across roughly 20,000 frames. Collaborated directly with surgeons to translate open-ended clinical questions into measurable ML metrics. The resulting architecture is now used as an objective surgical performance metric and won Penn's 2026 "Discovering the Future of AI" Award.
TA for BE 5140: Rehabilitation Robotics under Dr. Michelle Johnson. Assisted graduate students in translating research findings into the development of affordable therapeutic technologies, bridging the gap between clinical issues and engineering design, and delivered workshops on biomechanics, DFMA, forward kinematics, sensor integration, and rapid prototyping.
Conducted root-cause analysis for cable products using Lean Six Sigma techniques, including experimental design and statistical analysis. Led vendor testing, initiated development of a flexion performance testing device, and automated data collection for electrical measurements.
Developed and validated test methods for endoscopic medical devices. Contributed to prototype design, assembly, and troubleshooting while coordinating across departments on product development; submitted an invention disclosure to the Patent Review Board.