BaseballMindAI is an end-to-end multimodal pipeline for baseball broadcast analytics, integrating a domain-fine-tuned ASR module, an LLM slot extractor that resolves player nicknames without raw SQL generation, and a YOLO+OCR state-machine that reads scoreboard metadata for cross-modal event verification. Key events—home runs, strikeouts, walks—are auto-annotated as structured records; a natural language voice or text query interface returns matched clips on demand.

The core innovation is a three-layer multimodal pipeline—domain-fine-tuned ASR, LLM intent slot extraction, and a YOLO+OCR state-machine—eliminating manual clip annotation. The LLM layer resolves player nicknames and routes structured intents to backend APIs, preventing raw SQL generation. Temporal alignment of speech transcripts, subtitles, and scoreboard OCR output enables cross-modal event verification, reducing false positives in noisy broadcast environments.

The event-driven architecture targets three verticals: coaching teams gain rapid retrieval of situational clips for opponent scouting and player performance analysis; broadcast production teams automate highlight indexing and reduce post-production workload; fan engagement platforms enable natural language search without database expertise. The structured event schema—covering inning, count, base state, and pitcher-batter matchup—supports cross-season analytics and multi-sport portability.