Productized AI sports-coaching pipeline — video in, narrated coaching out

End-to-end video analysis pipeline on Google Cloud Run (8 vCPU, 32 GB RAM, no GPU): Gemini 2.5 Flash watches the footage and emits JSON segments with timestamps and coaching text; Google Cloud Video Intelligence (Person Detection) extracts 12 skeletal landmarks at 10 FPS; Gemini 2.5 Flash TTS narrates at 24 kHz, speed-adjusted via ffmpeg atempo; ffmpeg composes the final video (footage + skeleton overlay + caption + narration); results land in Firestore. Generalises to any video-analysis vertical (training, fitness, technique coaching).

Processing time runs ~2 min for a 30-second clip. 13 languages supported. The economics ($2.79/hr to analyze 5fps 1080p) clear the bar for consumer-subscription pricing against $250–400/yr competitor pricing. The architecture is replicable for any coach-replacement product (golf, tennis, martial arts, dance, music technique).

1. Stand up a Cloud Run service (8 vCPU, 32 GB RAM, no GPU is enough).
2. On upload, call Gemini 2.5 Flash with the video; parse output as JSON segments (timestamp ranges + coaching text).
3. In parallel, call Google Cloud Video Intelligence Person Detection for 12 skeletal landmarks at 10 FPS.
4. Send each coaching segment to Gemini 2.5 Flash TTS at 24 kHz; pad/compress audio length to segment length via ffmpeg atempo.
5. Compose final video with ffmpeg: original footage + skeleton overlay + captioned coaching text + narrated audio.
6. Write results (rendered video URL, segments, landmarks) to Firestore.
7. For 1-hour videos, fan out: split into ~12 chunks, run steps 2–4 in parallel, merge with a single text-merge prompt; wall-time collapses to one chunk's runtime.

Sequential single-pass processing on hour-long video burns wall time even though compute is cheap. The parallel-chunk fan-out (12 chunks → merge) is the unlock; without it the UX is "upload then come back tomorrow", which is fatal at consumer pricing.