Use a graph for connected questions
Represented extracted entities and relationships in Neo4j so retrieval could traverse scientific connections rather than return isolated text chunks.
Case study 02 / Applied AI + data
AstraNode
A citation-grounded GraphRAG explorer that connected hundreds of NASA bioscience publications under a 48-hour build constraint.
- Context
- NASA Space Apps Challenge
- Product
- Scientific literature explorer
- Stack
- Python, FastAPI, Neo4j, React, Gemini
- Result
- Phoenix winner / Global Judging
608NASA publications
48hBuild window
GraphRAGCitation-grounded retrieval
WinnerPhoenix challenge
01 / Problem
Scientific discovery was trapped across disconnected papers.
The challenge was to make relationships across NASA bioscience research inspectable without reducing the experience to opaque semantic search. Answers needed a visible route back to the source material.
The 48-hour constraint required a narrow product loop: ingest the corpus, extract useful relationships, build a graph, retrieve connected evidence, and expose citations in a usable interface.
02 / Architecture
Documents became entities, relationships, and cited answers.
Python services processed the corpus, Gemini and LLaMA supported entity and relationship extraction, Neo4j stored the graph, FastAPI exposed retrieval, and a React and D3 interface made the evidence explorable.
The retrieval layer preserved source references so users could inspect the publications supporting an answer.
03 / Decisions
The product was designed for evidence, not novelty.
Keep citations in the interaction
Made sources part of the answer experience so generated output remained inspectable.
Separate ingestion from exploration
Kept corpus processing and extraction distinct from the API and interface, allowing each layer to evolve independently.
Optimize for a demonstrable loop
Prioritized one end-to-end path that could be evaluated within the challenge window instead of spreading effort across disconnected features.
04 / Outcome
From 608 publications to a winning product in 48 hours.
AstraNode advanced from the Phoenix NASA Space Apps Challenge to Global Judging, demonstrating a complete retrieval product rather than a static research prototype.