Research

96.1% macro F1 on handball and 92.5% on basketball with sub-4 ms inference and 34.04 ms end-to-end pipeline latency.

Reduced post-event highlight processing time by 87.5% while maintaining 90% event-selection accuracy.

Released a synthetic labeled dataset with 365,383 sports frames covering seven standardized game situations.

Demonstrated practical NL-to-SQL interaction quality across four LLMs using human evaluation metrics.

Unified fragmented action-spotting methods into one reusable library pipeline for faster experimentation.

Broadcast editing style appears as the dominant factor affecting cross-domain action-spotting transfer.

98.6% classification accuracy with 4 ms inference and 34.04 ms end-to-end pipeline latency.