Really cool approach. The "Ollama for classical ML" framing makes it instantly clear what this does.

I've been building CLI-first tools myself and the pattern of wrapping complex workflows into simple terminal commands is underrated. Most devs I know would rather type one command than spin up a Jupyter notebook for a quick prediction.

Curious about the model format — do you plan to support a registry where people can publish pre-trained models, like Ollama's library? That would be the killer feature for adoption.