I'm building BeeBuzz for ML Zoomcamp 2026 and it's still rough. The idea: every bee species has a characteristic buzz. If a species goes quiet in an area, you could notice weeks before the flowers stay empty — and still have time to move hives and change the plan.
Right now it takes short lab buzzes and gives back species and confidence in seconds. I'm working with 200 Mendeley lab recordings in total (5 species), holding out 40 to check while I keep iterating: 80% accuracy on the holdout, with the weakest species (Centris cinza, F1 0.57) clearly needing more data.
Two things I want to be upfront about: confidence scores run overconfident (they're vote shares, not calibrated probabilities), and the buzz-type classifier scores 100% only because the filename prefix predicts the label — a recording-setup artifact, not biology. Lab recordings, no field noise yet.…