note.md

note.md is the project I started because my own research workflow was driving me up the wall. I'd read a PDF in one app, take notes in a second, manage citations in a third, and write in a fourth — re-deriving the same context every time I switched windows. The friction wasn't the writing. It was everything around it.

Then the cloud AI tools arrived and offered to make the friction disappear by doing the thinking for me: summarize this, draft that. But the deliberate thinking is the research. Outsourcing it doesn't speed you up, it hollows out the part that matters. So I set out to build the opposite tool.

What it is

note.md folds the whole research loop — reading, sourcing, note-taking, citing, and writing — into a single Mac app. It's Zotero + Obsidian + your PDF reader in one, built around a few stubborn principles:

• Local-first. Notes, sources, the search index, and every AI feature run on your machine. No claim, document, embedding, or model output leaves the device.
• AI as a librarian, not a ghostwriter. The AI surfaces passages you've already read and the evidence for and against your claims. It doesn't write your prose for you.
• You own your markdown. Everything is stored as plain markdown files you control, so you can walk away with your work at any time. The twist: you dont write it in markdown syntax, but rather in a block based editor. It feels like Confluence or Notion, but is acutally markdown.
• Deliberate craft over feature count. I'd rather ship a handful of features that are genuinely trustworthy than a long list that isn't.

Where the hard parts were

Most of the engineering depth lives in making "local-first AI" actually work on a consumer Mac — on-device PDF extraction, embeddings, hybrid retrieval, and a local LLM doing structured extraction, all fast and private enough to be usable. The articles below are my notes from building that: the problems I hit and how I solved them.

note.md is the debut product of ARSoftware UG, the company I founded in 2025.