Seon Min Kim
← All projects
SoloLocal LLM

BioGraph Intelligence

From a PubMed search to extracted claims, support and conflict checks, research gaps and a knowledge graph, running on a local LLM.

When
July 2026
Context
Personal project
Team
Solo
My role
Design and build
Tools
Python, Streamlit, Ollama (local LLM), PubMed, Pydantic, PyVis
Links
Live app, GitHub

What it does

BioGraph Intelligence runs a full evidence workflow on a local LLM: PubMed search, extraction of subject–relation–object claims, clustering evidence by theme, judging whether studies support or conflict, finding research gaps, and building a knowledge graph. It ends with a research brief and an evidence-grounded copilot.

What broke and what I changed

What happenedWhat I changed
Repeated calls to a paid LLM API used up the credits during development.Moved the LLM layer to Ollama, running locally.
With 20 papers in one prompt, the local model often returned prose instead of JSON.Extracted one paper at a time, with JSON validation and retries.
Streamlit reran the whole pipeline every time the copilot was used.Kept papers, claims, themes and briefs in st.session_state.

Limits and next

Extraction uses abstracts, not full text. Evidence scoring does not yet weight sample size or study design, and there is no MeSH or UMLS normalization yet.