A recruiting-intelligence layer for men's basketball.

A transparent talent-evaluation score that put players from different levels (a high-school recruit, a club standout, a college transfer) onto one common comparability scale so they could be viewed side by side, with cross-level translation treated as developing and unvalidated.

It was explicitly a comparability measure, not a prediction of future outcomes. That was a deliberate design call: it was scoped as a cross-level comparability measure rather than an outcome forecast, and shipped as an auditable, versioned score instead of a black-box "will succeed at the next level" number.

Webster was designed as a computer-vision engine to turn game film into structured, play-level evidence (proposing who did what and when where the footage supports it, and abstaining otherwise), plus a per-skill profile on each player card. It orchestrated well-understood open-source building blocks (object detection, multi-object tracking, scoreboard reading, jersey-number reading, pose, and appearance embeddings) into a single pipeline rather than leaning on one proprietary model.

It was abstain-first and human-in-the-loop by design: it proposed rather than asserted, reported an explicit confidence with every output, and abstained when uncertain instead of guessing. Results routed through a human correction step; those corrections were captured so they could inform model training once rights-cleared, labeled data was available.