Living AI

A Living AIis a real generative assistant packed into one AT-1 container of about 33 MB: a tiny distilled reader, an on-device embedder, and the index of your knowledge base, every tensor SHA-256-verified on load. It answers questions grounded in your documents, runs entirely offline, and declines anything it can't find in what you gave it. You never train a model; you point it at your content.

Build & ask from the CLI

# turn any folder of docs into a verified Living AI (~33 MB)
at1 living-ai build ./my-knowledge-base  atom.at1m

# ask it, grounded, offline, on-device
at1 living-ai ask  atom.at1m  "what is our refund policy?"
at1 living-ai ask  atom.at1m  "how many vacation days do we get?" -k 2

build scans a folder for text, Markdown, HTML and similar files, splits them into clean passages, embeds them, and packs everything with the on-device reader into a single verified .at1m. ask retrieves the most relevant passage and has the reader phrase a grounded answer; pass -k 2 to give it more context. Decoding and asking are free; only building (which compresses) is metered against your plan.

From the desktop app

In the desktop app, open Living AI, drag in a folder of documents, and click Build. You get a portable .at1m you can chat with in-app or share with a teammate, they only need the file, no install of your knowledge base required.

What's inside the file

• A distilled reader, a small instruction-tuned model that turns a retrieved passage plus your question into a natural-language answer.
• An on-device embedder, turns each question into a vector to find the right passage, with no network call.
• Your knowledge index, the embedded passages and their sources, so every answer can be traced back.
• Integrity, each component is SHA-256-verified as it loads; the container is an auditable artifact you can sign and store.

Honest limits

Atom is a small, private, verifiable assistant, not a frontier chatbot. It answers questions your knowledge base actually covers and declines the rest with an honest “I don't have that in this knowledge base.”Best results come from a real corpus (dozens of pages or more); a handful of sentences gives it little to retrieve from. It grounds every answer in what you packed in, it will not invent facts that aren't there.