PhysioLens

🎖️ Honorable Mention (Best Use of Bright Data) at Stanford TreeHacks 2026, top 5 of 200 projects. PhysioLens is an AI-powered physical therapy platform that turns any camera into a real-time rehab coach, using pose detection to count reps and correct form, and Claude to generate exercises and clinical summaries.

A couple of months before this hackathon, I had surgery, and my recovery ran on weekly physiotherapy: show up, get a new set of exercises, go home, and do them on my own until the next visit. That is where it quietly falls apart. The mistakes that slow healing are almost never dramatic, they are an elbow drifting forward on a curl, a shoulder rotating during a lateral raise, a back arching under a press, small deviations that feel close enough but stretch recovery out by weeks. You cannot feel them yourself, and between appointments there is nobody there to catch them. PhysioLens is the AI physical therapist I wish I had during that recovery: point a camera at yourself at home and get the kind of real-time correction you would normally only get standing next to a therapist. It earned an Honorable Mention for Best Use of Bright Data at Stanford TreeHacks 2026, landing in the top five out of more than two hundred projects. ### Turn on the camera, start healing For the patient, the whole thing is as simple as turning on the camera and starting to move. Google's MediaPipe tracks thirty-three points on your body in real time, and PhysioLens shows you a dual feed, your raw camera next to a live skeleton overlay, so you can see exactly what the AI sees. It counts your reps automatically from the angles at each joint, tracks which phase of the movement you are in, and the instant your form slips it says so out loud, things like elbow drifting, keep it stable or you are using momentum, slow down. A ten-second countdown with on-screen positioning guidance gets you framed before counting begins, so you are coached from the very first rep instead of finding out at your next appointment that you did the whole week wrong. ### The doctor just describes it On the clinician side, PhysioLens takes the busywork out of prescribing. A doctor types a plain-language description, something like tricep extensions for post-surgical shoulder rehab, and Claude generates the entire exercise configuration: the camera angle, the precise joint-angle thresholds used to count reps, step-by-step patient instructions, automatic form checks, and supporting references pulled straight from PubMed. What used to be a paper handout becomes a fully configured, research-backed exercise in anywhere from thirty seconds to a few minutes. The platform ships with a handful of professionally tuned exercises and lets doctors spin up unlimited custom ones on demand. ### It watches every rep, then explains what it saw PhysioLens does not just count, it analyzes. After each session it produces a performance score from 0 to 100, weighing reps completed against the form problems it detected and how severe they were. Its analysis engine watches for nine categories of biomechanical issues, tremor and instability, compensation patterns, speed variations, limited range of motion, asymmetry, fatigue, and more, chewing through thousands of pose frames far faster than real time. The best part for a busy clinician: click any flagged issue and the recording jumps straight to the exact moment it happened, so instead of trusting a patient's self-report, the doctor sees the evidence frame by frame. ### A physiotherapist that listens There is also a voice-driven Meeting Mode that runs quietly during a session. It transcribes continuously and watches for two things in particular: emergencies, so a phrase like my knee hurts really bad triggers an instant alert with an urgency score, and scheduling, so saying let us do a follow-up next Tuesday at three gets auto-captured onto the calendar. When the session ends, Claude writes a clean clinical summary covering chief complaint, patient mood, compliance, and follow-up recommendations, the kind of note that usually eats a therapist's evening. ### Grounded in real evidence Because this is healthcare, nothing gets recommended on a hunch, and this is where Bright Data earned the project its award. We started with Bright Data's Browser API, but static scraping only returned a couple of broken snippets per exercise. During a sponsor panel the Bright Data team pointed us toward their agentic framework with MCP tools, driven by the Claude Agent SDK, and that changed everything: Claude could now intelligently search, scrape, and extract across trusted clinical sources like NICE, NHS, and CSP, and we jumped from zero-to-two shaky references to nine to twelve high-quality, cited guidelines per exercise. A five-minute conversation with a sponsor beat hours of documentation, and it is the reason every recommendation in PhysioLens is backed by real clinical literature. ### Built in 36 hours All of this came together in under thirty-six hours on a React 18 and Vite frontend with a FastAPI backend. The build had its share of gremlins, the nastiest being MediaPipe's pose callback capturing stale React state through closures, which we fixed by leaning on refs instead of state for the values the callback needed live. We deliberately cut features like authentication and multi-patient support to keep the core experience flawless, because shipping one great thing beats shipping ten half-broken ones, and we actually finished a full six hours before the deadline with every core feature working end to end. PhysioLens was built with Aarav Matlala. ### The people I met along the way Honestly, one of the best parts of TreeHacks was the room itself. The opening ceremony had Garry Tan and Sam Altman on stage, and Sam came back later for a keynote fireside chat. Over the weekend I got to meet and talk with an incredible lineup of people I had mostly only read about: Walden Yan, co-founder and CPO of Cognition AI; Ali Partovi, founder of Neo; Quincy Larson, founder of freeCodeCamp; David Holz, founder of Midjourney; and Rajat Taneja, CTO of Visa. Building something I genuinely cared about, drawn straight from my own recovery, and getting to share it in a room like that is the part I will remember long after the code.

At a glance

Timeline: Feb 2026
Type: Hackathon Project
Event: Stanford TreeHacks 2026
Award: Honorable Mention - Best Use of Bright Data (Top 5 of 200)
Other Contributors: Aarav Matlala

Technologies

React 18 (Vite)
React Router
Google MediaPipe (Pose)
FastAPI
Python
Claude (Anthropic)
Claude Agent SDK
Bright Data Web MCP
PubMed API
Web Speech API
MediaRecorder API

Photos

Meeting the llama on the lawn at TreeHacks
In front of the TreeHacks 12 marquee
Talking through the project at TreeHacks
With Walden Yan, co-founder and CPO of Cognition AI, at TreeHacks
Sam Altman's keynote fireside chat at TreeHacks
With Ali Partovi, founder of Neo, at TreeHacks
With Quincy Larson, founder of freeCodeCamp, after his fireside chat
Selfie with David Holz, founder of Midjourney, at TreeHacks
With Rajat Taneja, CTO of Visa, at TreeHacks

Links

Machine-readable version of this page: https://www.sherwinvishesh.com/project/physiolens.md | site index: https://www.sherwinvishesh.com/llms.txt