The Gap Engine

what i made
its called the gap engine (anti-discovery engine). basically it pulls science papers off semantic scholar and openalex, builds a knowledge graph out of them with networkx, then hunts for structural gaps, aka holes between clusters of research that nobody has connected yet. an llm (groq) turns those holes into actual research questions, then it scores them by leverage (real betweenness centrality, not some fake made up number). theres also an inversion feature called the antimatter query, basically it checks "we know X causes Y, has anyone ever tested if Y causes X" and flags stuff thats never been tested. and the killer feature is historical validation mode, you run it on a 2005 paper corpus and it actually flags the network topology x epidemiology gap that later became a huge breakthrough in real life. proof it works in hindsight = proof it could work going forward.

built for the moonshot hackathon, deadline was june 30 2026.

what was hard
- the historical validation was straight up broken at first, backend and frontend data shapes didnt match and it was crashing
- had a bunch of fake numbers polluting the ui, like a leverage score that was literally just degree-sum labeled as "betweenness" when it wasnt, had to rip that out and make it real
- the demo graph kept surfacing boring gaps like "data mining to exome" instead of cool cross domain stuff, had to rebuild the graph generation to bias toward real cross domain voids
- node labels were straight up wrong sometimes (epidemiology showing as "political science" lol) so had to add a proper field normalizer
- the historical proof was originally rigged, it force constructed the answer instead of letting the detector find it organically. had to fix that so its actually earned, not staged

what im proud of
the historical mode is the big one. its not just a vibes based tool, it can point at a real graph from 2005 and say "yeah the engine would have caught this gap" and that gap turned into an actual scientific breakthrough (network science transforming disease modeling). thats the whole thesis of the project basically proven in one demo. also proud that i went back and ripped out the fake/inflated numbers instead of just shipping them, even though it wouldve made the demo look flashier.

how to test it
- backend: cd backend && python3.13 -m uvicorn main:app --port 8000 (gotta use python3.13 specifically, system python3 is 3.14 and breaks pydantic-core build)
- frontend: cd frontend && npm run dev (runs on localhost:5173)
- check out the historical validation mode first, thats the strongest proof point
- click into the antimatter query tab for the inversion stuff
- click on any gap to see the actual paper evidence backing it up (like "N papers touch A, M touch B, 0 touch both")
- theres a demo graph precomputed so you dont have to wait on live api calls to see it work