Agentic AI for Health
Overview
Agentic AI refers to systems where a large language model (LLM) acts not just as a question-answering endpoint, but as an autonomous agent — planning, calling tools, retrieving information, and completing multi-step tasks without constant human intervention. In healthcare, this paradigm opens the door to AI systems that can assist clinical workflows, automate research tasks, and synthesise evidence at scale.
Frameworks I work with
LangGraph
LangGraph models agent workflows as stateful directed graphs, enabling complex multi-step reasoning chains with loops, conditional branches, and human-in-the-loop checkpoints. This is particularly useful for clinical decision support pipelines that require evidence retrieval → reasoning → recommendation → validation cycles.
AutoGen & CrewAI
Multi-agent frameworks where specialised sub-agents collaborate on complex tasks. In biomedical research automation, one agent may retrieve papers, another extracts structured data, and a third synthesises findings into a report.
Retrieval-Augmented Generation (RAG)
RAG grounds LLM outputs in verified clinical evidence by retrieving relevant passages from a curated knowledge base before generation. This significantly reduces hallucination and is essential for any safety-critical medical application.
Use cases
- Clinical literature synthesis: Automated evidence summaries for systematic reviews
- Retinal report generation: Structured radiology-style report generation from ophthalmology images
- Research copilot: Automated paper ingestion, annotation, and insight extraction
Safety considerations
Agentic systems in healthcare require careful guardrails: output verification, human-in-the-loop approval for high-stakes actions, audit trails, and adversarial robustness. My research pays particular attention to evaluating failure modes before any clinical deployment.