N°01
2026
Reader · AI · Editorial UX
Marginalia.
A reading app that argues back. PDFs and long articles get a margin that pushes on the author's claims — not a chat sidebar.
01
Overview
Marginalia turns the margin of any text into a conversation partner. Upload a PDF, paste a URL, or pick an arXiv pre-print — Marginalia reads it, finds the load-bearing claims, and fills the gutter with objections, questions, and cross-references. The author is still the voice in the middle; the margin is the polite skeptic next to it.
The challenge
Readers swallow too much. A good editor leaves ink in the margin — sharp, specific, arguing — but most of us read alone. For technical and long-form writing, the absence of a pushback layer means every claim lands at full weight, whether or not it has earned it.
The solution
We built Marginalia around a single gesture: select a sentence, pull it into the margin, and a model returns a one-paragraph counter. It cites when it can, concedes when it must, and never volunteers a summary. The result reads like the margin of a careful editor's copy — not a chat interface wedged into a reader.
02
The Making-of
003 / 003
01
The brief, in one sentence
Give me a margin that reads like it's mad at me.
02
Against the summary button
Summaries are cheap. Summaries are also insulting — they assume you can't be bothered to read. Marginalia is the opposite bet: the reader is serious, the text is worth it, the margin is where the work happens. So we refused to ship a summary, and built the argument instead.
03
What we stopped doing
We killed the AI chat. We killed the 'explain this' button. We killed the right-rail abstract. What remains is a gutter, a reader, and a model that only speaks when called — and when it speaks, it disagrees first.
03
Outcomes
Marginalia launched quietly to a closed group of academic readers and essayists. Four months in, weekly retention settled at 64% — a number we had never hit on a reading product — and the average reader now leaves 18 margin-notes per long article, roughly four times what we measured on plain PDFs.
2,400
Weekly active readers
18
Margins per long article
64%
Week-4 retention
¢3
Cost per session
04
Colophon
Credits
Ayla Demir
Design Engineer
Niko Brandt
Chief Product Officer
Kenji Oshiro
Chief Technology Officer
Stack
Next.js 14
Electron
Claude Sonnet
PostgreSQL
pgvector
Tauri
Zustand
→
Next project
Case study · 2025
Halftone.
Read the case study