ANALYZ

HUMAN MOVEMENT. MACHINE PRECISION.

AI That Scores Human Movement From Video.

We decoupled physical kinematics from heavy pixels. Our specialized Action Quality Assessment model scores dance kinematics in 15 milliseconds — delivering full cloud video evaluation in 2–3s on GPU at 1/1,000th the compute cost of multimodal LLMs.

Full 30–60 FPS temporal kinematics. No 1 FPS downsampling loss.

15ms model scoring (< 2.5s full video on GPU).

Trained on proprietary commercial casting outcomes across thousands of gygs.

Analyz / Movement lab

Your video. Its motion. One timeline.

Upload an audition or analyse a sample. Track the process, then inspect body and face landmarks at any moment.

Rhythm specialistFull demo tracking / example scorecard
Awaiting tracking
Camera-space tracking
0.0 / 0.0 s
Head and face points
Live expression channels

Expression values appear when the face is tracked by the analyser.

Example scorecard
86.4/100

88.0th percentile

Shortlist signal
Rhythm sync92.0%
Line extension85.0%
Body coordination88.0%
Core control82.0%
Facial expression90.0%
Physical measurements
Beat offset: 42.0 ms
Peak extension: 162.0 deg
Arm difference: 18.0 deg
Torso tilt variation: 3.20 deg
Expression range: 0.380
Drop an MP4 or MOV to analyse

New uploads replace the preview and its tracking. The source and skeleton share playback controls.

Why Multimodal LLMs Fail at Human Movement

General vision models describe what they see. Analyz measures what actually happened, frame by frame.

Swipe horizontally on smaller screens to compare all three columns.

CAPABILITY

GENERAL MULTIMODAL LLMS (GEMINI / GPT-4o)

1,000x Faster • 1/1,000th Cost

OUR SPECIALIZED MOVEMENT MODEL

Frame sampling

1 FPS (Misses 29 out of 30 frames)

Full 30–60 FPS (Zero dropped frames)

Inference latency

35–50 seconds

15 ms (Model Inference) • ~2.5s (Full Video on GPU)

Cost per evaluation

~$0.05 to $0.15 per video

$0.00005 per video (1,000x cheaper)

On-Device Mobile

Impossible (Requires heavy cloud API)

Runs locally on iPhone / Android

Score calibration

Positivity bias (Clusters all dancers at 80)

Calibrated dynamic spread (50 to 98)

How It Works

01 / SENSE

Decoupled Physical Sensing

Decode video into 33-point 3D skeleton coordinates, audio drum onsets, and 52 facial blendshapes. Discard heavy pixels.

02 / MEASURE

Physical Kinematics Engine

Compute exact millisecond beat phase-locking, bilateral limb symmetry, and core balance stability.

03 / ASSESS

Action Quality Assessment (AQA) Network

Evaluates physical execution against real commercial hiring outcomes to output a calibrated quality score and population percentiles.

Developer API

A simple REST interface for commercial casting, performance platforms, and movement applications. Built to run on basic CPUs and phones.

REST / POST

curl -X POST https://cadence-backend-production.up.railway.app/v1/score_movement -F “video=@audition.mp4”

Illustrative request · API connection required for live scoring.

Stop guessing. Start measuring.

Human movement deserves more than a multimodal guess.

ANALYZ  ·  Movement intelligence. © 2026Privacy PolicyTerms of Service