Skip to content
All work

Graph-Native Retrieval for Clinical LLM Agents

Giving AI agents a way to follow the links between a patient’s FHIR records, and measuring whether it helps.

Role
Student researcher (solo)
Status
In progress
Course
EECS 4070 Directed Studies
Lab
York Data Mining Lab
Supervisor
Prof. Manos Papagelis
Timeline
Fall 2026
Stack
PythonFHIRLiteLLMpandas

What I’m building

  • A tool that follows the links between FHIR records in one call, instead of the agent fetching each linked record itself.
  • A test on FHIR‑AgentBench: the same agent, model, judge and questions, run with and without the tool.
  • As of September 2026: the benchmark is pinned and the data, 928,935 FHIR records for 100 patients, is checksum-verified. No results yet.

Next project

Barcelona’s passing network from the 5–0 Clásico: eleven player nodes joined by arrows, with Xavi the largest node, at the centre
Research

Tactical DNA

Does a coach leave a fingerprint on passing?