Portfolio

Rehabilitation Robotics · US Patent 11,311,447

A robotic rehabilitation system designed to improve agility and dynamic balance in patients with neurological and musculoskeletal conditions. The system uses real-time sensor feedback to automatically adapt challenge level to patient performance, keeping patients in the therapeutic zone — not too easy, not overloading.

Developed through university research and refined with clinical partners. The core idea: static rehab equipment is boring and sub-optimal; intelligent systems that respond to the patient in real time produce better outcomes.

Biomechanics Research · PNAS, 2021

A unified theoretical model that predicts both the paths people choose and the speeds they walk — using a single energy-optimality criterion. The kicker: prior models handled either path or speed independently. This one explains both simultaneously, from first principles.

Relevant beyond basic science: this kind of model underpins how autonomous systems (robots, prosthetics, exoskeletons) should plan motion to minimize effort and maximize safety. When the model is right, the robot moves like a human.

Gait Science · Gait & Posture, 2017

How does the human nervous system control stability when the environment is actively working against you? This study characterized how people adapt locomotor control in environments that either aid or resist stability, and what that tells us about fall risk and rehabilitation targets.

The findings are directly applicable to clinical settings: understanding how gait adapts to perturbation informs better treadmill protocols, better testing standards, and better interventions for fall-prone populations.

ML · Data Science · Sensor Fusion · 2020–Present

From wrist-worn motion sensors to full-body IMU arrays — building ML pipelines that extract meaningful biomechanical insights from noisy, real-world wearable data. Constrained computation, ambient noise, and realistic activity diversity are the actual problem. Lab data is easy.

The core challenge: making models that work when the user isn't cooperating — walking awkwardly, phone in a bag, smartwatch askew. Production ML for health and movement is harder than academic ML by an order of magnitude.