PUBLIC RESEARCH DEMO · Squat replays and AI model demonstrations use simulated data. No clinical, coaching, or real-person accuracy is established. Back to portfolio project ↗
MOTION INTELLIGENCE / RESEARCH 0.6
Move smarter. See every rep.
A connected movement-analysis workbench for back squats. Capture body landmarks, break down each repetition, and collect evidence for an exercise-specific model.
NO VIDEO UPLOAD · 33 POSE LANDMARKS · LOCAL SCORING
MOVEMENT PHASEECCENTRIC
YOUR WORKSPACE
Everything in sync.
01 EXERCISE ACTIVE
PIPELINE
From camera to insight
5-STAGE PROCESS
01CameraBrowser stream
→
02LandmarksMediaPipe
→
03SegmentsSquat cycles
→
04EvaluateRule baseline
→
05ReviewSave / label
i
Prototype, not a trained technique model
MediaPipe estimates pose landmarks. Scores currently come from transparent heuristics. Real ML evaluation requires consented, expertly labelled data and held-out testing.
01 / LIVE MOVEMENT STUDIO
Squat analysis.
Track repetitions and study the movement behind every score.
SAMPLE MODE
FF / MOTION ENGINE READY
◎Ready to analyseStart a sample session or enable your camera
Sample mode uses generated landmarks and a deterministic squat sequence. Scores are illustrative; no videos or models are downloaded.
MOVEMENT SIGNAL
Knee angle over time
DEMO STREAM
180°130°80°
SESSION TELEMETRY00:00
Technique baseline
—/ 100
ML pattern classifier: not loaded. Scores shown are heuristic.
Complete a repetition to see estimated movement quality.
⚑ RULE-BASED / NOT TRAINED
REPETITIONS0
KNEE ANGLE—
VISIBILITY—
PROCESSED0
LIVE COACHING CUE
✳
Position your camera to show hips, knees and ankles.
Repetition breakdown
0 REPS
Your completed reps will appear here.
02 / SESSION INTELLIGENCE
Session history.
Private session summaries saved on this device. No cloud connection.
TOTAL SESSIONS0
REPS TRACKED0
AVERAGE BASELINE—
LAST WORKOUT—
LOCAL DEVICE STORAGE
Saved workout sessions
ONLY SUMMARIES STORED
No sessions yet. Start a sample workout or use the camera, then select “Finish & save”.
↯
Storage and privacy
Session summaries are stored using browser localStorage. Clearing browser data can delete them. Export your records before clearing storage. No raw webcam video is saved by the website.
03 / LEARNED MOVEMENT PATTERNS
Model lab.
Train a separate movement-pattern model on reviewed repetitions, then test local inference in the browser.
CAMERA DATA NEEDED
YOUR TRAINING SET
Review & label repetitions
Complete a session to add examples here. Label them as an initial collection workflow. Labels should be reviewed against a documented rubric. Synthetic sessions are never eligible for real training.
No names or video recordings are exported. Label exports are never uploaded automatically. Generated samples are marked ineligible regardless of these boxes.
No examples yet. Visit Live coach and finish a squat session first.
CURRENT SCORING SYSTEM
Heuristic engine v0.1
Scores based on pose-derived knee-angle depth, estimated torso lean, tempo and tracking stability.
Pose detectionMediaPipe model, camera mode only
READY*
Rule-based evaluatorPrototype thresholds
READY
Labelled datasetNeeds consented diverse recordings
PLANNED
Trainable pattern classifierPython pipeline with participant-held-out evaluation
BUILT
Calibrated quality scoreMust be tested against coach labels
PLANNED
*Model files download from MediaPipe on first camera use. No session images are sent to a server.
CLASSIFIER / LOCAL INFERENCE
Train once. Evaluate every rep.
This website supports a genuine trained linear model, but the included example was fitted to synthetic simulations only. It must not evaluate real users.
No movement classifier loaded. Baseline scoring remains available.
Real training: export labelled camera repetitions → run Python trainer locally → import the resulting model. The classifier predicts a movement pattern, not a safety score.
!
Scientific limitations
A single camera cannot assess every aspect of 3D technique. Do not treat the 0–100 baseline as an injury or medical risk assessment. Label reliability and model calibration need validation.
DATASET RESEARCH / HUMAN MOVEMENT
Does movement over time matter?
NOT CAMERA VALIDATED
UI-PRMD deep-squat repetitions, with the study authors’ pretrained latent time-series. Compare summary statistics versus temporal patterns. This is a research experiment, not live-coach scoring.
Critical: Upstream author model was trained before participant splits, so both reported results have representation leakage. Neither result validates phone-camera tracking, weighted gym squats or technique safety.
OFFLINE RESEARCH · NO CAMERA REQUIRED
Five algorithms trained and compared on the same 180 recorded human squat repetitions. The best exploratory accuracy is not a deployable technique score. Compare model errors and participant variability below.
HUMAN REPETITIONS180
HELD-OUT SUBJECTS10
EXPLORATORY ACCURACY87.8%
MACRO F10.878
EXPERIMENTAL FEATURE SET
Means, standard deviations and six ordered windows of four latent channels.
CONFUSION MATRIX (ACTUAL × PREDICTED)
Non-optimalCorrectNon-optimal8010Correct1278
Compare all ten held-out participants
Research integrity: Data is from unloaded laboratory rehabilitation squats. Poor generalisation to a new person is still possible (the lowest fold reached just 55% in the temporal experiment). The trained model is never fed camera landmarks.
Original UI-PRMD publication · Source research data · Open Data Commons PDDL 1.0. This browser contains aggregate study results only, not participant recordings.
RESEARCH 0.6 / MULTI-MODEL EVALUATION
Five trained models. One test protocol.
RESEARCH ONLY
Every result is from participant-held-out testing on author-pretrained latent features. Select a model to inspect its confusion matrix and all ten participants.
SELECTED MODEL
Logistic · temporal features
EXPLORATORY ACCURACY87.8%
WORST PARTICIPANT55.0%
CONFUSION MATRIX: ACTUAL × PREDICTED
Non-optimalCorrectNon-optimal8010Correct1278
HELD-OUT SUBJECT RESULTS
Do not use these predictions on the camera: The upstream source autoencoder was fitted before the outer subject splits. Its latent features differ from browser landmarks. No clinical or technique-quality validation is claimed.
05 / MOTION REPLAY LAB
Replay every motion.
Rewind joint trajectories, inspect measured geometry and compare research-only sequence models. No camera needed.
Source: generated synthetic motion. The illustrated skeleton, phase and class are created for algorithm integration tests. Human squat quality and real-world accuracy are NOT established by these examples.
KINEMATIC PLAYER · SYNTHETIC
Body trajectory
OFFLINE
01 / 45
SEQUENCE RESEARCH WORKBENCH
Explore labelled motions
Knee angle—
Torso estimate—
Pose visibility—
Phase—
Replays recorded synthetic data from the included sequence fixtures; not a filmed human or a reconstructed UI-PRMD subject.
Both models were trained end-to-end on generated 45-frame motion sequences, using six virtual identities for training and two held-out virtual identities for testing. These are real model fits, but testing synthetic rules is not evidence of exercise-coaching quality.
Human dataset research remains separate: the older UI-PRMD accuracy numbers use author-pretrained latent embeddings and have upstream leakage. No claim that these neural networks outperform the human-data baselines, or that they can run on webcam pose sequences, is made.
04 / SYSTEM ARCHITECTURE
Everything connected.
One pipeline from visual evidence to movement intelligence and future model training.
01 / INPUTDevice camera or generated sampleVideo frames remain in the browser · no Flutter Android YUV conversion
03 / MOVEMENT ENGINERep segmenter + extracted featuresKnee angles, tempo, trunk angle and visibility gate
↓
04 / BASELINE SCORINGExplainable heuristic evaluator (v0.1)Not a trained or clinically validated technique score
↓
05 / DATA & FEEDBACKSession reports + human labelsBrowser storage and user-initiated JSON downloads
⋮
NEXT ITERATION / SEPARATE PIPELINEPython training → browser pattern inference → validationPortable classifier is implemented; real-world accuracy remains unverified
INTEGRATION WITH FLUTTER
Shared movement specification
The browser prototype mirrors FitnessFlex's existing squat state machine and feature extraction. The final implementation should define shared, tested data schemas so mobile and web use comparable feature and score versions.
SECURITY BY DESIGN
Zero service credentials
No OpenAI key, no Firebase config, and no backend credentials ship with this prototype. Rotate the previously exposed key and scrub its presence in Git history before using any connected service.