README

A kinder Kind of AI.

statuscoming soon languages10 small modelson your CPU organizercloud or local

Modules

Next-gen distributed AI. Instead of one giant model, a team of hyper-focused, feather-light models, each built and trained from scratch for a single job. They run on an ordinary CPU with every byte of memory accounted for, and answer in milliseconds.

Tulip

38M params

Spreadsheet importer

Hand Tulip a spreadsheet and she reads its preview, then writes a tiny TulipScript program that maps every row into the right table of your business database. That program never runs as raw code: a locked-down interpreter executes it and checks every instruction. If a sheet doesn't fit, Tulip refuses cleanly. No guessing, no silent damage.

Jasmine

~2–5M params

Design engine

The lightest petal. A few million parameters that compose your site's whole look in milliseconds, right on your CPU: palette, font pair, type scale, corners, density and the order of every section. Every choice comes from a vetted catalog, and code verifies it (contrast computed, fonts checked against your language) before you ever see it. She learns from the designs owners approve.

Daisy

27M params~300 ms per answer

Answers questions about your own data

Ask in plain English, like "which services earned most last quarter?". Daisy's 27 million parameters turn your question into a precise query spec in about 0.3 seconds on a single CPU core. Code checks the spec, compiles it to SQL in 0.4 ms and runs it read-only, then shows you the exact rows and the SQL behind the answer. No GPU needed, and when she doesn't know, she says so.

Principles

What a kinder Kind of AI means to us:

Small models, on your own computer. The precise work is done by small, fast models on an ordinary PC.
The organizer runs where you choose: on our server, or on your own PC.
Code checks everything. No model judges another model, and nothing a model writes is ever run as code.
Every value has a source: a line in your documents, a cell in your spreadsheet, or your own answer.
"No" is a normal answer. When a model isn't sure, it says so, and the question comes to you instead of a guess.
Private stays private. Revenue, staff numbers and client names are never published. Bank and card numbers, ID numbers and salaries are never even extracted.
You approve every public fact before anything is published.

Status

In development. The first release is coming soon.