← Back to demo hub · Biomechanics preview
Preview · runs locally on GPU

Biomechanics & Injury Risk

SportAI watches game and practice video frame-by-frame, turns each player into a skeleton, and flags movement patterns that signal fatigue, asymmetry, or injury risk — before a coach or trainer would spot them by eye.

What it does, in plain English

You point the pipeline at a game video. It tracks every player, measures how their body moves (knee flexion angles, hip drop, landing mechanics, stride symmetry, etc.), and writes a running log. If anything trends the wrong way — a player landing flatter on their right leg, a jumper whose hip drop keeps growing — the system raises an alert so staff can pull them or re-test.

The five stages

01
Ingest
Live camera or recorded MP4
02
Pose (YOLOv8)
17 keypoints per player, 30fps
03
Biomech calc
Joint angles, symmetry, fatigue
04
Anomaly detect
Compare to player baseline
05
Alert & log
Trainer-facing dashboard + history

Metrics tracked

MetricWhy it mattersTypical flag
Knee flexion angleLanding stiffness → ACL risk< 30° on landing
Hip drop (Trendelenburg)Core / glute fatigue> 5° asymmetry
Stride symmetry indexHamstring / ankle compensation> 8% L/R delta
Vertical jump deltaCumulative fatigue vs baseline> 15% drop over game
Landing force (proxy)Joint load, overuse riskSpike above P90

What an alert looks like

2026-03-04 · Q3 · 04:18
HIGH #24 Ramirez — right-knee flexion angle trending below baseline on last 6 landings (Δ −12°). Asymmetry index 0.14 > 0.08 threshold. Recommended: sub off for post-game re-assessment.
2026-03-04 · Q2 · 09:51
MED #11 Brooks — stride symmetry decaying (9.3% L/R delta). Possible left-hamstring guarding. Continue to monitor next 5 minutes.

The stack behind it

YOLOv8-pose

Ultralytics model, 17-keypoint skeletons at 30 fps on a single GPU.

TimescaleDB

Stores millions of raw frame-level keypoint rows, indexed by time and player.

Summary Postgres

Per-player baselines, alert history, rehab notes — the warm layer.

Prometheus + alerts

Ops-grade metrics and escalation so the pipeline itself is observable.

Auto-tuner

Thresholds are re-calibrated nightly from the player's own baseline.

Docker Compose

Full stack spins up in one command for testing on a spare box.

Why this is a preview and not a live demo. The biomech pipeline needs GPU compute, a real video source, and a player-baseline database to produce useful output. All three live on the founder's workstation today. When a client signs on, that client's team videos and baselines are provisioned into a dedicated tenant — so their data never leaves their own environment. This page exists so you can see what the tool is without having to install it.