Canis

Inspiration

Modern AI answers everything in the same confident voice, whether it knows the answer cold or is quietly guessing. That gap is where trust breaks. Anyone who lives with a dog can read them in a second. They never says a word, yet their ears, posture, and eyes tell you exactly how sure it is about what comes next. Canis started from one question: what if an AI wore its uncertainty on its face, the way a dog does?

What it does

Canis gives an on-device language model a dog avatar that reflects its real internal disposition in real time. As the model reasons, Canis reads its internal state (confidence, hesitation, conflict between competing answers) and renders it as living body language: ears up and alert when it is sure, a tilted head when it is weighing options, a lowered and wary posture when it is guessing. You stop reading walls of text to judge how much to trust it. You just look at the dog. Everything runs on the phone, so nothing you ask ever leaves your device.

How we plan to build it

The avatar is driven by a disposition layer, not by sentiment analysis of the output text. We will run an Apertus model fully on-device with MLX using a forward-only pass, and extract signals from the model's own internal state as it generates. Those signals will map into a small, interpretable disposition space, and each axis binds to a facial and postural expression on an Angular powered dog. To make the readout trustworthy, we built a disposition evaluation harness: a cross-axis rating study with Wilson confidence intervals and a three-rater inter-rater-reliability pass, so a given expression reliably corresponds to a given internal state rather than just looking cute.

Challenges we are expecting

Real on-device inference is the hard part. Getting Apertus running forward-only under MLX on an iPhone, inside memory and thermal limits, will take lots of engineering. We have done an analysis of the handful of legible dog expressions without misrepresenting what the model is actually doing. Then is the important UI aspect of getting an avatar that users will enjoy interacting with

What's next

We plan to make a first version of the ios app available over Test Flight for initial test users, and gather feedback. Also test the model dispositions in several scenarios