Research is mostly reading. The tools were mostly filing.

Academic workflows never got the software everyone else did.

A literature review means holding forty papers in your head at once. In practice that looked like thirty browser tabs, a folder of PDFs, a separate notes doc, and no way to ask a question of the material you'd already read. Search was keyword-based, so you found papers containing your words rather than papers answering your question. And nothing was shareable — collaboration meant emailing PDFs.

The market had tools for pieces of this. None of them held the whole workflow.

Competitive setSemantic Scholar, Research Rabbit, Elicit, Scite. Each strong on one step, none on the sequence.

Twenty interviews before a single screen.

I ran 20+ interviews with researchers across disciplines and career stages. The pattern that mattered wasn't a missing feature — it was that every tool assumed research is a search problem, when researchers experience it as a comprehension problem. They don't struggle to find papers. They struggle to hold what the papers said.

That reframe set the roadmap. I ran a competitive analysis to find where a contextual relevance engine could differentiate, then defined and sequenced the initial feature set against it.

• Contextual searchask a research question, not a keyword string
• Summaries and TLDRsdecide whether to read before reading
• Chat with the paperinterrogate a source directly
• Literature review generatorsynthesis with human curation
• Bundling and shared commentscollaboration on the actual sources
• Data extraction and visualisationpull the numbers out

I wrote the PRDs for each and storyboarded the MVP flows.

Use an interface people already know.

Researchers were already fluent in one AI interface pattern — the centred search box that turns into a conversation. Every one of them used ChatGPT or Claude daily. Inventing a novel paradigm would have meant spending their attention on learning our product instead of on their work.

So the landing screen looks deliberately familiar: one input, centred, nothing else. The intelligence is underneath, not on the surface.

The information architecture followed researchers' own mental models rather than the data model — papers group the way a person's filing system groups them, not the way a database does.

Trust is a design material in academic tools.

Researchers prioritise accuracy over speed — the opposite of the consumer patterns most AI products borrow. Interfaces that optimise for a fast answer read as untrustworthy here. Citation tracking, source attribution, and visible user control weren't compliance features; they were what made the product usable at all.

The second lesson was about pacing information. Academic content resists summarising, so progressive disclosure did the work instead — comprehensive underneath, digestible on the surface, with the researcher deciding how far to open it.