Context
At NCSA, I worked on improving reasoning and retrieval in biomedical AI systems by incorporating knowledge graphs into LLM workflows. The research focused on building a structured knowledge base of genes, phenotypes, and biomedical relationships to support causal question-answering over CRISPR perturbation data.
What I Built
I built a Neo4j knowledge graph with 100K+ biomedical triples by harmonizing Monarch ontology data, creating a structured layer for multi-hop reasoning over biomedical entities.
I then developed a ReAct agent using LangGraph and DeepSeek-R1 to traverse the graph and support reasoning between genetic perturbations and phenotypic outcomes.
To expand the graph with new structured relationships, I fine-tuned HuggingFace REBEL on BioRED for biomedical relation extraction, achieving 0.64 F1 on BioASQ-style evaluation.
Challenges
The main challenge was connecting unstructured biomedical information with structured reasoning. Biomedical entities and relationships are noisy, domain-specific, and spread across ontologies and literature, so the system needed a structured representation that LLM workflows could reason over more reliably.
Impact & Takeaways
This experience strengthened my work in knowledge graphs, biomedical NLP, relation extraction, RAG-style reasoning, and agentic AI systems for scientific question-answering.