What I built
Built a read-only sports research dashboard and Elo evaluation pipeline. The saved 1,693-game held-out test reports 63.94% straight-up accuracy and a probability error score of 0.221, where lower is better.
Data and model roles
Read-only source feeds → cached research dashboard → sourced context and probability comparisons. The recorded Elo evaluation uses separate training, validation and held-out periods; serving inputs differ from the backtest.
Implementation scope and verification
Implemented read-only research dashboard, cached context adapters and an Elo model with a stored train/validation/test evaluation. Optional Grok context adapter; current configuration unverified.
Quantitative evaluation, source-backed probability comparisons, caching and read-only API boundaries. Historical game-prediction accuracy is not a trading result.
About the evidence
Based on source review and saved project artifacts. Recorded measurements apply to the stated model, dataset and configuration; they are not claims of a current production deployment.