Schrollar
An AI-powered research-discovery platform that turns natural-language queries into curated academic-paper feeds. Shipped end-to-end in a 24-hour hackathon — placed 3rd of 22 teams and won PHP 10K at GDG Loyola Hackfest 2026.
- Next.js
- TypeScript
- Python
- Tailwind CSS
Problem
Researchers — especially students new to a field — spend hours filtering paper databases trying to phrase the right query. Discovery tools are tuned for exact keyword search, not the messy way real research questions get asked in plain language.
Approach
Schrollar accepts a natural-language research query and returns a curated, ranked feed of relevant papers. Queries are mapped to active research streams so a user describing what they care about gets back a feed shaped around that interest, not a literal keyword match. We shipped the full pipeline — query understanding, paper aggregation, relevance ranking, and the reading interface — end-to-end during a 24-hour hackathon.
Technical Decisions
Natural-language → research-stream mapping
Rather than embedding every query against the entire paper corpus, queries are first mapped to a smaller set of active research streams. Streams are aggregated lazily and reused across users. This kept latency tolerable and the cost predictable inside a hackathon budget.
24-hour scope discipline
We deliberately cut features that didn't directly support the demo path: query → ranked feed → readable result. Auth, persistence, and personalization were pushed past v1. The discipline is what got us shippable in 24 hours and on stage at Loyola Hackfest 2026.
Tradeoffs
Stream-first architecture vs. open-ended search
Mapping queries to a curated set of streams trades expressiveness for relevance and latency. It works well for early-stage exploration; users with very specific paper-level queries are better served by a keyword tool.
Hackathon shipping vs. production hardening
No retry/backoff, minimal error states, and best-effort caching. Good enough for a demo; the next iteration would put these in.
What I learned
- 01
Shipping a working data product under a hard 24-hour deadline forces brutally honest scope.
- 02
Mapping fuzzy human queries onto a structured taxonomy is often more useful than open-ended embedding search.
- 03
Demo polish is engineering work — judges scoring 22 teams notice the small things.