joygoLive AMC
AMC · AI Motion Coaching

One phone. Your form, judged as you move.

Stand your phone up and start the set. AMC reads your joints from the camera, judges each rep against a trainer's criteria adjusted to your own body, and gives a spoken cue the moment a correction is needed: no wearables, no straps, no video sent anywhere.

See how it works What this does not claim
A service by joygoLive, built here and operated here. For community sports, built with trainers.
Exercises judged today
Squat Romanian deadlift Overhead press
How it works

Five steps, no phone interaction

There is no setup procedure. You are not asked to measure a distance or a tripod height, only to stand where your whole body fits in the frame.

Stand the phone up
2-second baseline
Do the set
One spoken cue
Report

Place it anywhere reasonable

No measured distance, no tripod height, no gyroscope. The only question that matters is whether your body is in the frame, and if it is not, AMC reports that instead of guessing.

Calibrated to your own body

Two seconds of standing still sets your personal baseline. Depth and range are then read relative to you, so a long femur or a short torso is not mistaken for bad form.

One cue at a time

Mid-set you get a single spoken correction: the one that matters most right now. A list of six corrections cannot be acted on while you are under load.

Approach

Readable joint-angle rules

AMC does not compare your rep to a reference video. It measures joint angles, normalizes them against your own baseline, and applies thresholds a trainer agreed to. The rest of the product follows from that design decision.

Body shape is not an error

Template matching flags anyone whose proportions or tempo differ from the reference. Angle rules against your own baseline remove that false positive at the source.

Feedback you can act on

“Your knees are 12° narrower than your baseline” tells you what to move. A similarity score of 0.73 does not; it cannot be disputed, verified, or coached against.

Light enough to stay on the device

Angle arithmetic is cheap, so judgment runs on-device. Nothing is uploaded for a verdict: a privacy property and a latency property at once.

Labeled reps verify the rules

Labeled reps exist to measure the rules: the miss rate on real errors and the false-positive rate on good reps. The rules are the product, and the labeled data verifies them.

Three severity levels

Each error carries one of three severity levels, and the strongest correction is capped per error type by joint-load risk: the instruction never pushes you harder than the movement safely allows.

Reports from templates, no language model

Report sentences come from templates tied to what was measured. No model call means no server, no API key, and no chance of a fluent sentence that the numbers do not support.

What you get

The set, in numbers you can compare next week

Everything below is recorded per rep, so a set is not a single score but a per-rep series you can review.

Rep countCounted only when the rep was actually judgeable.
Range of motionHow deep you went, as a percentage of your own baseline.
TempoDescent, hold at the bottom, and ascent, per rep.
Errors and severityError type with one of three severity levels, plus how to fix it.

A report that reaches you where you are standing

When a set ends you are two meters from the phone, not holding it. So the result arrives in three layers instead of one screen you would have to walk over to read.

  • Spoken confirmation when the set ends.
  • Four numbers on screen, sized to be read from two meters away.
  • The full breakdown when you pick the phone back up: depth per rep, the rep-by-rep trajectory, tempo, one correction to work on, and the rest listed below it.
  • Coverage: which checks the camera angle allowed, which it did not, and where to stand to get them.
Set summary: the four large numbers
Reps10
Clean form8
Avg. depth94%
Set time48s

Illustrative figures: a real set fills these from your own reps.

Limits

Reps it could not see are excluded and explained

The easiest way to look accurate is to count everything and omit the frames that could not be read. AMC does the opposite: when a check cannot be made, it reports which one and why.

Lower body obscuredLegs hidden behind equipment or cropped by furniture.
Key joints out of frameA joint the rule depends on has left the picture.
Detection confidence too lowLighting, clothing, or distance pushed detection confidence below the threshold.
Standing at an angleNeither a clean front nor a clean side view, so the angle is unreliable.
Also stated explicitly
  • Tempo is measured and shown, but not judged. Calling a rep “too fast” requires a normal distribution across many people agreed with trainers, and that does not exist yet, so the numbers are displayed and no verdict is attached.
  • Only a short list of exercises is judged. Each one needs its own rules and its own field validation, so the list grows slowly.
  • A single camera cannot see everything at once. Front-only checks and side-only checks are named as unverified when you are not standing that way, rather than assumed to have passed.
  • AMC is still being tuned in the field with community sports organizations, not a public release. Thresholds are still moving as real-device data comes in.
Privacy

The video stays on the device

Judgment happens on your device, so the video never has to leave it. Exactly what is and is not transmitted:

Never collected Video, images of your face, your name, and your contact details. None of it is transmitted, and none of it is stored on a server.
Sent during field tuning While thresholds are being calibrated, numbers are sent automatically: rep counts, angles, joint coordinate values, and device information. The app states this on its first screen.
After tuning ends Once the thresholds are settled, coordinate transmission is switched off and returns to opt-in. The judgment itself never needed the network.

No personal details are requested: the participant number the app uses is issued by the app itself.