Tesseract
A local-first knowledge system that runs its entire AI pipeline on-device. No cloud calls, ever.
6
Models on device, zero cloud calls
31
Architectural decision records written
~55.6k
Lines tracked across Rust, TS and SQL
Why I built it
For years my capture system was a WhatsApp chat with myself. It never asked me anything, which is why it worked, and it had no structure at all, which is why finding something two months later meant scrolling past hundreds of messages.
Capture asks nothing
Every notes app asks where a thing goes at the moment you are trying to save it, which is the worst possible time to ask. Here there is one destination. Press Sync later and it groups what is there, names the groups, and lays them out.
One store, four lenses
Canvas, folders, timeline and search are four renderings of the same Collections, not four features. Change something in one and the others agree immediately, because there is one write path rather than four views doing their own updates.
The AI can't move anything on its own
Sync is the only process allowed to change what a Collection holds or what tags an item carries. Insights runs in the background and can only write to a feed that waits for approval. Move something by hand and it is never re-filed.
Writing, and getting your notes in
A rich-text editor with headings, tables, inline images and drawings, where wikilinks become real edges in the graph. Point it at an Obsidian vault and it imports notes, tags, folders and embeds, showing you what it proposed before anything settles.
