Every modality had its moment. Movement is next.

The pattern repeats across a decade of machine learning: stop hand-crafting features, train one model on the raw signal at scale, and every downstream task improves at once. Text went through it. Images went through it. Protein structure went through it.

Human movement has not. The raw signal exists — skeleton sequences, dozens of joints, twenty-five frames per second — yet the field still runs on hand-picked metrics: distances, speeds, counts. That gap is the company.

TextGPT2018Next-token prediction over raw text
ProteinsAlphaFold2020Structure predicted from sequence
ImagesDINO2021Self-supervised vision without labels
MovementWithoutBallopenSelf-supervised representation of how the body movesPrototype
Foundation-model moments by modality. The movement row is the one still open — our model is a working prototype, not a finished claim.

We measure how the body moves — not just where it goes.

Tracking systems answer where: positions, distances, sprints. The skeleton answers how: stride mechanics, deceleration control, left–right asymmetry, coordination under fatigue. That second layer is where form changes, habits show, and risk accumulates — and it is invisible in a dot on a plane.

WithoutBall learns this layer directly from raw skeleton sequences — self-supervised, no hand-labeled events. What it unlocks is catalogued under capabilities; how it is built is documented in the docs.

Where — a dot on a plane

How — 21 joints in motion

Same second of play, two signals. The left one is solved; the right one is ours.

Two channels into one knowledgebase.

Everything the model computes — embeddings, derived features, links back to the underlying movement segments — accumulates in a knowledgebase. Two kinds of clients sit on top of it.

Channel 01

Direct product

Coaches, performance staff, and clinicians work in our own interfaces — reports, review tools, monitoring views.

Explore solutions

Channel 02

Agent integration

External AI agents query the same knowledgebase on behalf of their users and answer with evidence.

The AI-agent platform
One movement knowledgebase

Football first. Not football-only.

The model learns bodies, not pitches. Nothing in its architecture assumes a sport — a skeleton sequence is the same object whether it sprints, strides, or lands. Football is the first vertical because that is where our data and domain depth are; the ones after it are already mapped.

Beyond Football