Code Icon
Case Study

Discovery showed clinicians needed biomarker interpretation—not storage

Overview

During discovery we shadowed clinicians across intake, lab review, and patient consultations. The loudest pain wasn't storing results—it was interpreting biomarkers into clear, empathetic guidance.

We built a multi‑step agent powered by LangGraph, LangChain, pgvector, and CopilotKit that translates raw results into plain‑language explanations and actionable next steps while maintaining context across sessions.

Architecture & Design
  • LangGraph Orchestration: Multi-step reasoning with modular nodes for data retrieval, memory recall, context building, recommendation drafting, and summarization.
  • pgvector for Memory & Context: Vector embeddings in Postgres enabled fast retrieval and continuity across sessions.
  • CopilotKit Integration: Real-time streaming between UI prompts, API calls, and agent reasoning.
Prompt Engineering Strategy
  • Persona-Guided Responses: Empathetic and scientifically grounded tone.
  • Layered Instructions: Step-by-step scaffolding from values to actions.
  • User-Friendly Output: Conversational, precise, and free of jargon.
Privacy & Security
  • AWS RDS, HIPAA-ready: Encryption at rest and in transit.
  • Isolated infrastructure: Compliance with client security standards.
  • Minimized retention: Essential data only; anonymized vector embeddings.
User Testing & Iteration
  • Clarity of Advice: Guidance remained understandable and unambiguous.
  • Tone of Communication: Supportive and guided, not overwhelming.
  • Actionability: Recommendations led to concrete next steps.
Results
  • Speed to Market: Prototype in under two weeks; production shortly after.
  • Personalized Experience: Contextual, empathetic advice tied to biomarkers.
  • Scalable Architecture: LangGraph + pgvector scales to new data sources.
Key Takeaways
  • LangGraph excels when reasoning paths must be explicit.
  • pgvector provides a scalable backbone for personalization.
  • Persona-grounded prompt engineering makes complex data approachable.
  • Live user testing drives trust, usability, and accuracy.