About the project
Inspiration
Training alone means being your own observer. You finish a set, but it can be difficult to tell whether your posture changed, how far you leaned back, or which repetition deserves another look.
That was the starting point for Motion X Ray. I wanted to turn a workout recording into something more useful than a video sitting in my camera roll: a way to inspect movement, compare it with references and understand what to focus on next.
Milo, our training companion, brings a friendlier personality to that experience. The aim is to make reviewing a set approachable for someone who does not have a personal trainer beside them.
What it does
Motion X Ray turns a short workout video into a movement report with replayable evidence.
The prototype supports workflows for lat pulldowns, dumbbell curls, bodyweight squats and incline dumbbell presses. Users select an exercise, upload their set and review detected movement cycles, available measurements and experimental comparisons with reference videos. Segment replay helps them return to the actual moment behind a result.
We distinguish between what the system measured, what it could not observe and what remains provisional. Reference resemblance is not presented as a definitive technique grade, and visibility is not treated as proof of accuracy.
How we built it
The application combines a Next.js and TypeScript frontend, a FastAPI backend, and a containerised Python worker using MediaPipe Pose Landmarker.
The processing workflow is:
Upload → pose estimation → movement segmentation → geometric measurements → reference comparison → report and replay
We researched exercise guidance, consolidated the relevant movement criteria and used reference videos to establish initial comparison baselines. These remain experimental and need further review and calibration.
A central design decision was to keep measurements separate from language models. Deterministic code calculates the movement metrics and reference comparisons.
For coaching, NVIDIA Nemotron through Nebius Token Factory has a deliberately bounded role: selecting from evidence-linked coaching candidates. Local code validates the selection before displaying the corresponding guidance. The model does not calculate angles, invent scores or assess raw workout footage.
We also used Nebius coding workers for implementation and review, with bounded requests and usage tracking. That development workflow is separate from the app’s runtime coaching integration.
Challenges we faced
The hardest problem was deciding when a measurement deserves to be shown.
Camera perspective, occlusion and brief tracking gaps can interrupt an otherwise visible movement. We had to preserve partial cycles and unavailable measurements rather than turn uncertain observations into confident results.
Processing speed was another challenge. We tested smaller analysis inputs and frame-rate changes, then compared their effects on processing time and detected movement. This helped us understand that a faster pipeline is only useful if it preserves enough information for the assessment.
Reference calibration also required care. Different demonstrations vary in viewpoint, pace and execution. We therefore kept the initial comparisons provisional instead of presenting one reference as a universal definition of correct form.
The final recorded walkthrough also exposed a coaching limitation: the runtime request returned no validated coaching response. The app preserved the measured report and replay rather than inventing a substitute answer.
What we learned
We learned that an AI fitness product needs to make its uncertainty understandable.
A useful result is more than a number. It should explain what was observed, connect back to the video and make missing evidence visible. We also found that language models are more accountable when their role is narrow and their output can be checked against measurements.
What comes next
Our next priorities are stronger reference calibration, more reliable coaching validation and clearer exercise-specific feedback.
Beyond that, we want to connect set reviews with workout history and routines, expand the exercise library and explore an optional compact camera for recording sets. These are future directions, not completed capabilities.
Motion X Ray is an early personal-training prototype, not a medical device or a replacement for professional assessment. Its purpose is to help people look more closely at their movement and make their next set more informed.
Log in or sign up for Devpost to join the conversation.