Show HN: ride-recap, teaching a LLM my taste to automate cycling highlights
10 hours ago
- #Automated Video Editing
- #LLM Applications
- #Cycling Technology
- The author built an automated pipeline called ride-recap to create cycling highlight reels from GoPro footage and Garmin .fit files using Gemini LLM.
- The pipeline involves steps like ingesting data, detecting interesting segments via Gemini vision, telemetry, Strava API, fusing clips, and outputting videos for different platforms.
- Teaching LLMs 'taste' required iterative prompt improvements based on ground truth data, such as adjusting for power zones, negative examples like traffic jams, and positive landmarks.
- Candidate selection avoids repetitiveness by using a narrative pass and greedy re-ranking with crowding penalties to spread clips evenly across the ride.
- The project, driven by personal passion, highlights lessons in video processing, LLM calibration, and the importance of iteration for developing automated systems.