Vanderbilt explores AI practice-video tagging; effectiveness remains untested

The customer problem is delayed practice feedback. Engage the coaching lead, athletics analytics, student analysts and campus IT. Ask how much footage awaits review, which decisions require faster feedback, and whether existing tools already meet the need. A bounded engagement could establish a tagged-video baseline and test one event category. The value hypothesis is reduced review delay at acceptable error rates, subject to measurement. Discovery should include analysts' availability, camera access and the cost of correcting false labels. This account supports a scoping conversation only; do not promise performance gains, staff reductions or same-day delivery from an untested pipeline.

Fit is assisted video annotation with evidence attached to each proposed event. Map camera exports to a protected processing queue and review interface; confirm timestamps, frame quality, consent and retention before model access. Compare a simple detector with a general multimodal model on held-out practices. Measure event precision and recall, player-attribution errors, processing latency and reviewer correction time against manual work. Test difficult lighting, occlusion and similar clothing. Choose local, cloud or hybrid processing only after data and cost review. Autonomous agents have limited initial relevance; use bounded jobs with no ability to alter training plans or disclose clips.

Start with a data inventory, access approval, annotation rubric and representative evaluation set. The athletics analytics lead owns operations, while coaches adjudicate disputed labels and IT manages access. Train analysts in quality review and retain enough staffing for corrections. Governance checkpoints should precede footage transfer, expanded event coverage and operational use. Proposed acceptance criteria are lower median review turnaround than the measured manual baseline, no unreviewed player attribution in coach reports, and complete source-clip traceability. Record user adoption through accepted versus corrected annotations. Camera integration, insufficient labels and model drift could erase apparent efficiency; preserve manual review as a fallback.