Embedding Space
Scalable Player Profiling
Embed every player’s movement signature into a common space. Scouting becomes similarity search: “find me players who move like this one.”
What it is
One window describes a moment; a distribution of windows describes a player. Player profiling aggregates thousands of movement signatures per athlete into a stable profile — a location and a shape in embedding space that captures how that player moves.
Because every profile lives in the same space, comparison stops being an exercise in adjectives. 'Explosive', 'economical', 'press-resistant' become measurable neighborhoods — and 'find me players who move like this one' becomes a query with a ranked answer.
How it works
Profiles are running aggregates over a player's windows: mean signature, spread, and drift over time. The mathematics on top is deliberately simple — the power sits in the representation underneath. Profiles sharpen as the corpus grows: at proof-of-concept scale the model separates players within a match; robust fingerprinting across seasons is a data-scale milestone, not an algorithmic leap.
What you get
For scouting, profiling compresses hours of video review into a shortlist worth watching — every candidate arrives with the movement evidence that put them there. For development staff, it turns 'they look sharper this month' into a measured statement against the player's own history.
| Parameter | Value | Notes |
|---|---|---|
| Profile | distribution over signatures | Mean, spread, and trajectory in embedding space. |
| Similarity query | k nearest profiles | A ranked shortlist; each hit with per-window evidence. |
| Development view | profile drift over time | The same player, measured against their own past. |
| Scale today | within-match separation | Validated at proof-of-concept scale; see the status note below. |
Status & roadmap
Prototype. Similarity search over movement windows and embeddings runs in the internal workbench; the scouting workflow described here is the product framing of that capability. Claims scale with the corpus — and are labeled accordingly.