Two models from a privacy-first canine health platform, running on your device.
Nothing leaves your device. Both models run in this browser tab —
there is no server here to send a photo or a recording to. That is Dally's actual
architecture, not a demo shortcut.
Model loads on first use (~8 MB).
≈83% test accuracy on 4 balanced classes (MobileNetV2 fine-tuned
on ~4,000 images) — a research prototype, not a clinical instrument. Roughly one
photo in six is misread, and it will be confidently wrong on unusual angles,
breeds and lighting.
Model loads on first use (~90 MB, cached afterwards).
General-purpose AudioSet filtered to its dog-vocalization classes. It hears
bark; it does not know why your dog is barking.