birdbird

1 min read

Aside left

Screenshot from birdbird: highlights video player, with a list of detected species beside it linking to the best sighting of each Screenshot from birdbird: bar chart of bird species detected from video, each with confidence % Screenshot from birdbird: bar chart of bird vocalisations detected from audio, each with confidence % Diagram from birdbird: how it works, showing the parallel video and audio processing pipelines Screenshot from birdbird: credits and acknowledgments page listing models, frameworks, and licenses used

What?

See it in action at birdbird. Detects bird species from both the audio and video of motion-captured bird feeder clips, then publishes the results as an interactive site - a highlights reel that bookmarks the best sighting of each species, so one click jumps straight to it, plus browsable charts of every detection and its confidence.

Why?

Sharpen my AI-assisted development skills. Provided close direction, requirements clarity, quality oversight, and design decisions - leveraging Claude Code for implementation while maintaining full project vision and technical governance. More detail in birdbird - Human Contribution Summary.

Tech

  • FFmpeg for general processing of input clips
  • ML inference using publicly available pre-trained models:
    • Trim input clips to keep segments with birds. YOLOv8 + COCO dataset.
    • Identifies bird vocalisations using BirdNET
    • Identifies bird visuals using BioClip (optionally, process on remote GPU)
  • Deploys on Cloudflare Workers and R2.

See birdbird credits for full list of tech and dependencies.