See exactly what JumpOnion analyzes — start to finish, on one page.
First, a quick look inside the app. Then a complete 2Lo (double loop) diagnosis from a real practice session — the same engine that runs on every customer upload. Look through it, then decide if it's worth $49/month to your skater.
Faces mosaicked for privacy. Recorded inside the JumpOnion app during a real practice session.
Computer vision measures the jump.
Below are the metrics from one real 2Lo upload. Pose estimation tracks the skater across all major joints; biomechanics turns those points into measurements that match what a coach would see — but exact, repeatable, frame-by-frame.
Not AI-generated. Computed with computer vision + biomechanics. Same numbers every time you upload the same video.
Rule engine measures. LLM translates.
A deterministic rule engine compares every metric to clean-jump targets and decides severity. Then an LLM turns that decision into language a parent and a 9-year-old can both act on. The LLM never invents numbers — it's an interpreter, not a judge.
This jump looks good overall. No major issues were identified.
The skater lacks the vertical impulse needed to achieve sufficient air time for a double jump. Combined with a slightly delayed arm pull, the skater runs out of time to complete the rotations in the air, resulting in an under-rotated landing.
The green "excellent" is the main verdict. The red "also noted" is where most tools stop. That's where JumpOnion starts.
Every claim is traced to a measurement.
Three evidence points support the "also noted" warnings. Not hallucination. Not guessing. If you want to challenge a claim, click through to the metric.
The diagnosis becomes a sequence.
Four off-ice drill categories, picked from the proprietary drill library, ordered by when they should happen relative to the next ice session. Each drill explains why— so you know what you're fixing, not just what to do.
Drill names shown here are generalized for public display — paying customers see the full named drills, exact sets, and target bands inside the app.
Vertical Impulse Drill
3 × 8Builds consistent, explosive vertical impulse — trains the body to deliver a precise, repeatable launch for maximum jump height.
Air-Position Proprioception Drill
3 × 8Forces instant 'freeze-frames' in various jump positions, sharpening proprioception so the skater can find and hold a tight axis in the air.
Arm-Snap Speed Drill
3 × 15Targets tightening speed directly. Train the arms to snap in the instant the blade leaves the ice, shortening the time-to-tight window.
Centrifugal Resistance Drill
3 × 12Builds explosive 'braking' power against centrifugal force. Essential for holding a super-tight air position once rotation is loaded.
What the skater feels. What the parent watches.
Most AI tools forget that the people executing the fix are a 9-year-old on the ice and a parent in the stands. JumpOnion gives each of them their own page.
“The instant your blade leaves the ice, imagine your body becoming a tightly wound spring — push up, then instantly snap into a spinning top.”
Stand at the side, slightly behind the takeoff point. Watch the takeoff knee: does it fully extend upwards, like a powerful piston, or is there hesitation? Immediately after takeoff, watch the arms: do they snap to the chest in one explosive motion, or is there a visible delay?
Want this for your skater?
Every JumpOnion subscription includes unlimited jump uploads, full access to the 54-drill library, historical tracking across sessions, before/after comparison, and email support from a skating mom (me).
I tested it. A general-purpose LLM misclassified a 2Lz as a Toe Loop and called an imperfect air position "perfect". LLMs hallucinate; rule engines don't. JumpOnion uses the LLM only as a translator on top of a deterministic biomechanical engine — that's what makes the diagnosis trustworthy. Read the full breakdown on the blog.
DM @cikibuilds on X or @jumponion on 小红书. I read everything.
The diagnosis, metrics, and training plan above come from a real production task (0dc3eb30, March 30, 2026). Not a mockup. Not a demo. The same engine ran on this upload as runs on every customer upload.