Watcher uses motion detection to trigger its camera pipeline. When movement is detected, the system captures a burst of images, selects the sharpest frame, locates the animal, and attempts to classify its species. Camera detections that cannot be classified confidently may be recorded as unidentified animals. A separate continuous audio process listens for bird sounds and identifies only confident, named bird species. Camera and audio observations are stored in a shared SQLite database. The system compiles daily statistics and wildlife summaries, which can be viewed through a local web interface and delivered through Telegram. All camera, audio, database, web, and inference functions run locally on a Raspberry Pi 5. The processes coordinate their memory usage so that the language model and wildlife inference do not exceed the device’s available resources.