PitchScanner
Place the ball, the attackers and the defenders on a pitch, and PitchScanner searches StatsBomb's open 360 data, about 1.3 million frames from 426 matches, for the closest real situations. Then it shows what those moments led to: how often they produced a shot, how much xG they were worth, how often possession was lost.
- Built with
- Python, FastAPI, Postgres with pgvector, React, TypeScript
- Status
- Live
What it does
- Every outcome comes with a confidence interval and the baseline for that part of the pitch, so a claim that a shape is dangerous can be checked.
- Filters are written in plain language, like “Bayern at home, last fifteen minutes, trailing”. Rules parse them first; a model is only a fallback.
- Situations are encoded three ways (grid, polar and tactical) and searched with pgvector.
- The query path is built for a 150 ms p95.