Show almost any enterprise a useful AI workflow and the demo goes well right up until someone from security asks the only question that matters:where does the data go?

For every mainstream assistant the answer is some version of “to a datacentre we operate, under a policy you should read.” For a lot of teams that’s fine. For the ones I work with — regulated industries, client-confidential material, an internal audit function with opinions — it ends the conversation. Not because the tool is bad, but because the paperwork to say yes costs more than the productivity it buys.

So I built the other thing.OnDevice.ai is a private AI workspace that runs on your own machine: streaming chat, project-based context, file analysis, and real PowerPoint, Excel and Word generation. Inference runs locally through Ollama or LM Studio. Cloud is an option, not an assumption.

The core architectural bet is that document reliability comes fromschema-validated specs and fixed generators — never from asking a model to emit binary Office files. The model produces a JSON spec; a deterministic generator produces the bytes; structural validation hard-gates the download. A file that downloads is a file that opens.

And the honest part: on raw capability, Claude and ChatGPT win, and it isn’t close. A frontier model has more knowledge, better reasoning and far better taste. What this has instead is ahard floor — the deck won’t win a design award, but it will open every time, cost nothing per document, and never leave the laptop. The ceiling is set by the model and rises as local models improve; the floor is set by the pipeline and doesn’t move.

The write-up covers why it exists, what it is, the “specs before bytes” bet, the architecture and trust boundary, the hybrid GGUF/MLX model policy, how web grounding stays off unless it’s needed, a feature-by-feature comparison with the commercial assistants, where it genuinely earns its place — and the list of capabilities it deliberately refuses to build.