A multimodal pipeline for assessing beehive health without opening the hive, exposed as a live Fetch.ai uAgent fleet with a chat interface and a real-time dashboard. Built with a team.

The guiding principle was honest evaluation. Every acoustic model is validated hive-held-out, meaning no colony ever appears in both the training and test folds. The score therefore measures generalisation to colonies the model has never heard, which is the only thing a deployed system actually faces.

What works, and what doesn’t

Signal	Modality	Honest performance	Verdict
Colony strength	Acoustic	0.646 hive-held-out	Works — matches the published baseline
Input gate (bee/no-bee)	Acoustic	0.58 hive-held-out	Weak; only 6 colonies available
Queenless	Acoustic	0.18 hive-held-out / 0.88 within-hive	Confounded — see below
Varroa mites	Acoustic	Not learnable	Label is single-class in this data
Varroa mites	Vision	0.986 accuracy	Usable

Population is the one solid acoustic model.

The queenless result is the interesting failure. It scores 0.88 within-hive and collapses to 0.18 hive-held-out — a textbook case of a model learning which colony it is listening to rather than queen state, because each colony in the dataset is only ever recorded in one state. A single random train/test split would have hidden this completely and produced a number that looked great.

Varroa cannot be detected acoustically from this data at all: across roughly 100 inspections, none reach the 3% economic treatment threshold, so the label has only one class. The trainer refuses to fabricate one. Mites stay a vision task.

Why report the bad numbers

Because reporting both within-hive and hive-held-out scores is what exposes identity leakage. Publishing only the optimistic number would have made the project look considerably better and been worthless to anyone trying to build on it.

The apiary feed is simulated — there are no live instrumented hives behind this — but the models and the agent decision logic are real.