Learn to build sophisticated agentic workflows using LangGraph for clinical documentation automation in this comprehensive conference talk. Explore the engineering principles behind multi-agent systems through a practical clinical scribe use case that transforms messy medical conversations into structured SOAP notes using a four-agent pipeline. Understand why breaking complex AI tasks into specialized agents improves accuracy, auditability, and debuggability compared to monolithic prompt approaches. Master LangGraph's core concepts including state machines, nodes, edges, routers, and cycles while implementing a shared MedicalScribeState for audit logging and testability. Dive deep into each agent's responsibilities: transcription processing, structured data extraction with validation retries, negation handling, temporal qualifiers, and deterministic normalization. Discover techniques for reducing hallucinations through structured inputs, no-new-facts prompts, and low-temperature settings, plus implement consistency checking to catch omissions, contradictions, and invented facts. Follow an end-to-end demonstration covering the complete pipeline from audio transcription through extraction, SOAP note generation, and consistency validation. Explore advanced topics including retrieval grounding for mapping diagnoses to ICD-10/SNOMED codes, quality measurement strategies per agent, and production considerations for PHI handling, redaction, prompt injection protection, monitoring, and human review workflows.

Syllabus

Intro: Orchestrating Agent State Machines for Clinical Documentation
What “Agentic” Means Engineering View: Nodes, State, and Routing
The Clinical Scribe Use Case + 4-Agent Pipeline Overview SOAP
Why This Matters: Documentation Burden, Burnout, and System Impact
Why Not One Big Prompt: Accuracy, Auditability, and Debuggable Workflows
Why Clinical Notes Are Hard: Messy Dialog, Negation, Time, Uncertainty
What We’re Building: Inputs/Outputs, Pipeline, and Two Key Design Choices
Why LangGraph: State Machines for Branching, Retries, and Traceability
LangGraph Vocabulary Crash Course: State, Nodes, Edges, Routers, Cycles
The Shared MedicalScribeState: Audit Logs, Errors, and Testability
Schemas & Structured Extraction: Pydantic Models as an API Contract
Node Implementation Template + Transcription Agent Responsibilities
Extraction Agent Deep Dive: Validation Retries and Negation Handling
Negation, temporal qualifiers & deterministic normalization in extraction
From structured encounter to SOAP note: summarization constraints & editable sections
Reducing hallucinations: structured inputs, no-new-facts prompts & low temperature
Consistency agent: catching omissions, contradictions & invented facts
Consistency checks taxonomy: hard rules vs semantic judging vs safety scans
Graph orchestration with LangGraph: explicit routing, safe failures & debugging
End-to-end demo walkthrough: transcription → extraction → SOAP → consistency
Retrieval grounding: mapping diagnoses to ICD-10/SNOMED as a pipeline node
Measuring quality per agent: datasets, metrics & iteration strategy
Production readiness: PHI, redaction, prompt injection, monitoring & human review
Setup, key takeaways & next steps feedback loop, better grounding