UNDER THE HOOD

How the AI Judges Every Rep

REPPED's scoring isn't a black box or a marketing exaggeration — it's a real computer-vision pipeline running in your browser. Here's exactly what it does, in plain language.

1. Pose estimation: mapping your body to a skeleton

The moment your camera turns on, an on-device pose-estimation model (Google's MediaPipe Pose Landmarker) maps your body to 33 key points — shoulders, elbows, wrists, hips, knees, ankles, and more — many times per second, directly from your video feed. This runs as WebAssembly inside your browser, using your device's GPU where available. Nothing is uploaded to a server for this step.

2. Per-exercise analysis

Each exercise has its own analyzer that watches for the specific joint angles and body positions that define a valid rep:

Push-ups — elbow bend depth, chest proximity to the floor, and full lockout at the top.
Squats — hip-to-knee depth (breaking parallel) and standing all the way up between reps.
Burpees — a complete stand → floor → stand cycle; partial cycles don't count.
Plank — hip alignment held in a straight line for the full duration, not rep count.

A rep that stops short of full depth, or an elbow that doesn't fully lock out, is measured against your own body proportions in the frame — not a fixed camera-angle-dependent number — and is flagged as incomplete rather than silently counted.

3. Visibility and effort tracking

The system continuously checks how much of your body is visible and in-frame. If visibility drops too low for too long, the analyzer can't reliably judge your reps and the match reflects that. Separately, it tracks motion energy across key joints throughout the 30 seconds — being in active motion for a healthy share of the battle is what full effort credit is based on. Standing still for most of the clock and bursting for the last three seconds scores worse than steady, consistent work, exactly as you'd expect from a fair judge.

4. The final score

Your rep count (or hold time for plank), form quality, pacing, and effort combine into a single score from 1 to 10 for that match. Hitting the target rep count for an exercise with clean form is what a perfect 10 represents — the target varies by exercise since a burpee and a push-up demand very different paces.

5. What happens when the camera can't see you well

Poor lighting, an unusual camera angle, or stepping out of frame all degrade what the model can measure. Rather than guessing, REPPED's analyzers are built to flag low-visibility periods explicitly, and a lightweight motion-based fallback engine takes over automatically on older devices or browsers where the full pose model can't load — so a match never silently breaks, it degrades gracefully.

6. Where this runs, and why that matters for privacy

All of the above happens on your device. Your camera frames and body-position data never leave your browser for scoring — only the final numeric result (score, rep count, exercise) is ever sent to our servers, and only for signed-in players who want their Elo saved. During a live 1v1 match, video streams peer-to-peer directly to your opponent; our servers relay connection signaling only. Full details are in our Privacy Policy.

Honest limitations

This is camera-based pose estimation, not a certified biomechanics lab. It isn't a medical device, it can be thrown off by extreme lighting or camera angles, and — like any client-side system — a determined bad actor can attempt to interfere with it. We don't claim otherwise. What we do build in: plausibility checks on submitted results, rate limiting, and a reporting system for suspected abuse. See our Anti-Cheat and Fair Play pages for exactly how that works, and what to do if you think an opponent gamed the system.