Nexus

Inspiration

We wanted to build something closer to how researchers actually think: starting from a rough idea, following connections, comparing directions, and gradually refining what matters. Nexus was inspired by the need for a research assistant that does not just retrieve papers, but helps map a field.

What it does

Nexus turns a natural-language research intent into an interactive scientific map. The user describes their topic, methods, and field; Nexus generates editable Semantic Scholar queries, retrieves seed papers, lets the user select what is relevant, and then expands the search using recommendations. From those papers, it builds a weighted graph based on embeddings, citations, co-citations, and user feedback. The system also computes a user profile vector that can guide future exploration.

How we built it

We built a Python backend connected to the Semantic Scholar API. It generates queries, fetches seed papers, calls the recommendations endpoint with positive and negative paper IDs, retrieves SPECTER embeddings, and builds a graph with weighted edges. The frontend is a standalone interactive demo using Three.js to visualize papers as a 3D research landscape. For the pitch, we also support precomputed real data so the demo stays fast and reliable without depending on live API latency.

Challenges we ran into

The biggest challenge was turning a broad research idea into a structured pipeline that still feels natural to the user. We also had to balance live API calls with demo reliability, handle missing embeddings or incomplete metadata, and avoid creating a graph that becomes visually unreadable. Another challenge was designing the weighting system so the graph reflects both semantic similarity and citation structure, while still leaving room for user interaction to shape the result. And that we started from scratch during this Hackaton.

Accomplishments that we're proud of

We are proud that Nexus connects the full loop from onboarding to real paper discovery, recommendations, graph construction, and profile vector generation. The demo uses real Semantic Scholar data rather than static fake content, while still being stable enough for a live presentation. We are also proud of the visual experience: instead of showing papers as a list, Nexus makes research feel exploratory, spatial, and interactive.

What we learned

We learned how powerful paper embeddings and citation signals become when they are combined instead of treated separately. We also learned that user feedback is essential: the same topic can lead to very different research paths depending on what the user select

What's next for Nexus

Next, we want to complete the RAG agent layer so users can chat with the graph and ask for research advice, missing concepts, or next-step queries. We would like to have a full on recommendation system without cascade models. We also want to integrate the possibility to extract a context from the user vector profile so Nexus can take one promising concept and have agents help to work on it. Other next steps include persistent user profiles, better graph clustering, better frontend/backend integration, paper summaries, and collaborative research maps for teams.