TikTok for researchers
The problem
Research discovery is noisy in a very specific way: people are not just looking for more papers, posts, and reports. They are looking for the few things that change what they know. Generic feeds tend to reward popularity or recency, which often means researchers see familiar material again while missing the insight that would have mattered.
Who it was for
The product was for people who need to keep up with fast-moving domains and cannot afford to read everything. Their current routine was a patchwork of alerts, saved searches, newsletters, and manual scanning. The break point was not access to information; it was deciding what deserved attention.
What we built
We built a feed that treats novelty as personal. Instead of asking only whether an item is important in the abstract, the system estimates whether it is likely to be new for a specific reader at their current level of expertise.
That product choice shaped the whole experience: onboarding had to capture a reader’s interests without becoming a survey, ranking had to balance freshness with relevance, and summaries had to explain why an item was worth opening.
My role & the technical calls
I owned the product direction and built the first backend. The early roadmap focused on the loop that mattered most: ingest content, retrieve the strongest supporting context, summarize it, and learn from reader behavior.
The RAG pipeline was designed to keep generated output tied to source material. I prioritized retrieval quality, source attribution, and evaluation of whether a recommendation was both relevant and non-obvious for the user.
The hard tradeoffs
The main tradeoff was breadth versus trust. A wider source universe made the feed feel alive, but weak ingestion or shallow summaries would quickly erode confidence. We kept the first version narrow enough to evaluate quality and learn from user behavior.
Cold start was the other hard edge. Too much onboarding would slow people down; too little would make the feed generic. The solution was to start with light preference capture and let the system adapt as users saved, skipped, and opened items.
Outcome & what I learned
The company was selected by Alchemist Accelerator with proposed funding. The project sharpened my view that AI products win only when the model behavior is attached to a painful user loop. In this case, the loop was not “search better”; it was “help me notice what I would otherwise miss.”