Bird Buddy and similar commercial smart bird feeders sell for $200-400 and rely on cloud subscriptions for species identification. The hardware is honest — a Pi-class compute with a decent camera and a microphone — but the recurring fee per identification ($5-10/month) is what funds the unit cost. A DIY version pushes the species model to run on-device and skips the subscription entirely.
The classifier doing the work is YOLOv8-nano fine-tuned on a regional bird dataset. There are pre-trained weights available for North America (Macaulay Library data), Europe (eBird), and a few national datasets in Asia and Australia. Fine-tuning to your specific yard's birds takes about 200 photos per species and a free GPU evening on Colab. After that the model runs at 6-9 FPS on a Pi 4 (1.5W draw), or 25-30 FPS on a Pi 5 with a Coral USB accelerator if you want continuous identification rather than triggered.
The reward mechanism — extra seed when a rare bird shows up — is what makes this fun rather than just another camera feeder. The feeder is two compartments: a normal one that's always open, and a 'reward' one with a servo-controlled door that opens for 30 seconds when the classifier sees a species you've marked as rare in your config. Birds learn the pattern within a few visits and start hanging out longer when the reward door opens.