Keel

private AI research assistant. fully browser-based. built for people who needed a system that was architecturally private, not just promised to be.

everyone else's “private AI” is a promise -- a company asking you to trust it won't look at what you send its servers. keel removes the thing you'd have to trust. the model runs on the gpu already in your machine, your documents never leave it, and there's no server on the other end to log them, leak them, or train on them. it uses the compute you already own and takes none of your data in return.

and what runs there is not a toy. ask it something hard and it breaks the question apart, gathers evidence across your documents, critiques its own draft, chases the gaps it finds, and stops when it stops learning -- a real research loop, multi-round and self-terminating, not one retrieval and a paragraph. retrieval is hybrid, bm25 for the exact term and semantic for the meaning with a cross-encoder ranking what survives, and if the gpu runs out of memory mid-answer it recovers and resumes instead of dropping the tab.

the private part is easy to describe. the harder part, the part that separates a product from a demo, is that a system this capable runs end to end in a browser tab at all -- a full model and an agentic research loop, on a consumer gpu, where nothing close to it was running before.