Full project description
Animals don’t communicate problems clearly, and by the time something is obvious, it’s often already serious. This subnet is designed to pick up early signals that something is off, using behavioural data rather than waiting for visible symptoms. Miners analyse short video clips, audio, and basic environmental context to identify patterns that indicate stress, pain, or abnormal behaviour. Instead of just classifying what’s happening, they assign a risk level and flag potential issues. The goal isn’t diagnosis, it’s early detection. Validators score outputs based on consistency, accuracy against labelled datasets, and agreement across miners. Known datasets (e.g. vet-reviewed cases or controlled scenarios) can be used to benchmark performance, alongside consensus scoring to identify strong signals. The initial scope would focus on domestic cats. It’s a controlled environment, behaviour is nuanced, and there’s a real gap in early detection. From there, it can expand into broader pet use cases, shelters, and eventually livestock monitoring. This isn’t about “AI for pets” in a gimmicky sense. It’s about building a reliable signal layer for animal well-being that can surface issues earlier than humans typically would.
Why it works on Bittensor
This works well as a subnet because it relies on pattern recognition across many independent models, where consensus and accuracy matter more than a single “best” answer. Different miners will pick up different behavioural signals, and the network naturally rewards those that consistently identify real issues. It also benefits from continuous improvement over time as more labelled data and edge cases are introduced. The competitive structure encourages better detection of subtle signals, which is where current tools fall short. Centralised models can do this to an extent, but they don’t incentivise diverse approaches or reward long-term accuracy in the same way.
