Archive

About these identifications

Species: Made by a local AI model and may contain errors. Identifications are reviewed and corrected over time.

Behaviors: Some are reliable (bathing, drinking are clear from video). Others like preening or vocalizing are often guesses — the model fills in what a species typically does at a birdbath rather than what it actually observed.

Sex and age: Easy calls like adult male House Finch are reliable. Female vs juvenile distinctions and most other sex/age calls are rough approximations.

How Birdwatch works

Hardware: A small camera detects motion and records short clips. An AMD GPU handles inference locally via ROCm.

Infrastructure: Everything runs in Podman containers on a self-hosted Linux server. Ollama runs the qwen2.5vl:7b vision language model locally. No cloud inference.

Pipeline: camera clip lands in an ingress folder. A systemd path watcher detects it and triggers processing. The coordinator extracts 5 frames and sends them to Ollama for identification. Results are written to JSON, the clip moves to archive, and a reporter script reads all JSONs, applies corrections, and generates the HTML report. A scheduled timer pushes the report to GitHub Pages.

Corrections and accuracy over time: Clips are reviewed via a LAN-only correction UI. When a bird ID is corrected, it gets saved and triggers automatic description regeneration and report rebuild. Before analyzing any new clip, the model receives a list of every species previously verified at this fountain. This should reduce repeated mistakes on birds the model has already gotten wrong once — but it is a nudge, not a guarantee.