Built and maintained a HIPAA

Built and maintained a HIPAA compliant production medical AI scribe system that uses multi-step LLM agents to extract clinical entities directly from physician-patient conversations and turn them into structured medical notes, cutting down the manual transcription work clinicians used to do after every visit.

Multi-stage clinical workflow: Documentation, coding, and review each have different logic and conditional paths, and the system needed to branch correctly between them without losing context.

Clinical accuracy: Generated notes had to be grounded in real patient history and clinical guidelines, not just plausible-sounding text.

Production reliability: As a live system handling real conversations, every agent run needed to be observable, debuggable, and monitored for cost and latency in real time.

Non-technical requirements gathering: Clinical needs had to be captured accurately from stakeholders without a technical background and translated into precise agent logic.

Approach:

Stateful agent orchestration with LangGraph Designed LangGraph-based agent workflows with conditional branching, allowing the system to move correctly across documentation, coding, and review stages based on conversation content.

Context-grounded note generation with RAG Built a RAG pipeline on AWS Bedrock with embeddings, so every generated note is grounded in the patient's actual history and relevant clinical guidelines rather than generic output.

Full production observability Integrated Langfuse across all agent runs to track token usage, latency, and model KPIs, giving the team visibility into system health and cost in production, not just at build time.

Clinical stakeholder collaboration Ran requirements sessions directly with clinical staff, converting their documentation needs into concrete agent behavior specs and validation criteria.

Results & Impact:

~40% reduction in manual transcription time for clinicians using the system.

Clinically grounded output, with notes tied to real patient history and guidelines instead of unsupported generation.

Full production observability, with token usage, latency, and model performance tracked continuously.

A workflow clinicians could trust, built through direct collaboration rather than a black-box handoff.

Provided Services & Solutions:

📌 AI Agent Development (LangGraph) 📌 RAG Pipeline Development (AWS Bedrock) 📌 LLM Observability (Langfuse) 📌 Cloud Infrastructure (AWS Lambda, S3, DynamoDB) 📌 Stakeholder Requirements Translation 📌 Production ML Systems

Tech Stack

Python · LangChain · LangGraph · AWS (Bedrock, Lambda, S3, DynamoDB) · Langfuse · TypeScript · REST APIs