PaperDive

Overview

PaperDive is a PDF reader with a study guide attached. You upload a paper and it opens in an ordinary reader on one side. On the other, the paper has been split into its sections, and each section can be summarized at one of three reading levels: Simple, Standard, or Technical. Select a phrase you don’t know and it opens a concept card. The card has terms of its own that you can select, and so on, as far down as you need to go.

Suppression of neuroinflammatory response to platinum wire implanted in mouse brain by monoclonal antibody SIF001 against Galectin-3

Sun JH, et al. · 2025 · abstract, shortened

Implanted electrodes lose signal over time, in part because of the immune response around them. A thin platinum wire stood in for an implant in the mouse brain, and mice were dosed twice a week for four weeks with an antibody against Galectin-3. Treated mice had about 60% fewer activated microglia around the wire (p < 0.0001), and their time on the Rotarod rose from 20 to 113 seconds (p < 0.01). The results suggest that blocking Galectin-3 may reduce neuroinflammation around neural implants.

SimpleStandardTechnical

this section, in short

A platinum wire modelled an implant in the mouse brain. An antibody against Galectin-3, given twice a week for four weeks, was associated with about 60% fewer activated microglia around the wire (p < 0.0001) and a longer time on the Rotarod, from 20 to 113 seconds (p < 0.01).

select a highlighted phrase

There is also a chat that has read the whole paper, a diagram for each study’s methods, and a pen and highlighter for writing on the pages.

It was built first for students and early-career researchers reading biology, neuroscience, bioengineering, and medical papers.

The problem

Papers are written for people who already speak the field’s language. A first-year reading one for the first time hits an unfamiliar term every other sentence, and has to choose between losing the thread and leaving the page to look something up.

The idea was to keep the real paper in front of you and put the help beside it, at the level you need, rather than handing you someone else’s summary in its place.

What I did

I’m building PaperDive, from the product and the design to the code.

How it works

A PDF is checked on the way in, by type, by size (25 MB at most), and by the %PDF- signature at the start of the file, and then stored privately. The server pulls out the text and page count, and a heuristic looks for the usual scientific headings to divide the paper into sections.Heading-based detection is simple on purpose. It is also where the rough edges are; see the end. Nothing has been summarized yet.

That is deliberate. A summary is generated only when you open a section at a reading level, and then it is cached, so the second visit is instant and costs nothing. The chat, the concept cards, the methods diagrams, and the educational media (Wikimedia Commons and YouTube, with attribution) follow the same rule: generate when asked, and keep what you can.

Under the hood it is Next.js, TypeScript and Tailwind on the front; Supabase for sign-in, the database and file storage; PDF.js for rendering; pdf-parse for extraction; Gemini and OpenAI for the language work; and Mermaid for the methods diagrams.

Decisions worth explaining

Generate late, cache everything. The beta has monthly limits per account, currently two uploads and fifty AI requests. Generating on demand and caching keeps cost predictable inside limits like those, and keeps the interface from feeling like one long loading screen.

Leave the references out of the chat. The reference list stays navigable in the reader, but it is excluded from what the chat reads.

Show text, not decoration. The media layer looks for diagrams, animations and videos for a concept, and each candidate is reviewed. If nothing accurate passes, the card falls back to plain text rather than a pretty picture that is subtly wrong.

Private by default. Papers belong to the account that uploaded them. Every protected route checks the user and the paper’s owner, database rows are locked down with row-level security, and the PDF itself is served through signed URLs.

What came out of it

The honest list of what doesn’t work yet is longer than the list of what does, which is a fine place for a first version to be.

• Text extraction can read two-column layouts, equations, tables and scanned pages out of order.
• Section detection depends on headings, so unusually formatted papers split badly.
• The reader shows one page at a time; continuous scroll isn’t built.
• Extraction runs inside a web request instead of a background job, so very large PDFs can hit hosting limits.

Next on the list: quizzes and flashcards for active recall, structured notes you can export, imports from PubMed, arXiv and Zotero, and a native iPad app built around the Apple Pencil.