Case StudyArchitect & AI Engineer

MedDiagnose AI.

A multi-agent medical diagnosis platform: seven AI pipeline stages and five specialist agents turn symptoms — typed, spoken, or photographed — into a structured differential diagnosis with safety checks.

Client
Independent Product
Year
2026
Tools
Claude API, OpenAI API, Gemini API, FastAPI, Python, Vue.js, PostgreSQL, Redis, Docker, Stripe

Overview

MedDiagnose AI is an exploration of multi-agent AI applied to one of the hardest interaction problems: helping a patient understand their symptoms. Users describe what they're experiencing through chat, voice, annotated photos, or an interactive body map, and a pipeline of cooperating AI agents produces a structured differential diagnosis, treatment suggestions, contraindication checks, and a plain-English report with PDF export. It is an educational and demonstration system, carrying prominent disclaimers — a design study in safe, structured medical AI rather than a clinical device.

The Agent Pipeline

Diagnosis runs through seven stages: a triage agent structures the intake, then diagnostician and research agents work in parallel, a specialist stage routes to one of five domain sub-agents (cardiology, neurology, dermatology, gastroenterology, psychiatry), followed by treatment, safety, and finally an empathy agent that translates clinical findings into humane language. Agents consult one another over an async message bus, drug-interaction and medical-image agents contribute where relevant, and a self-learning quality auditor reviews outputs. Progress streams live to the browser over server-sent events.

Multi-Provider AI

The LLM layer is provider-agnostic: Claude is the primary and recommended engine, with OpenAI and Gemini as alternates and Whisper handling voice transcription. This made the system a working testbed for comparing model behavior on identical structured-reasoning tasks.

Engineering Highlights

  • FastAPI backend with SQLAlchemy 2, Alembic migrations, PostgreSQL, Redis, and rate limiting
  • Vue 3 + Vite frontend with Capacitor mobile/PWA packaging and 12-language i18n
  • Stripe subscriptions, a 16-page admin dashboard, developer API portal, and white-label theming
  • Test coverage across pytest, Vitest, and Playwright end-to-end suites; Docker Compose deployment

Outcome

A complete, safety-conscious reference architecture for multi-agent AI in a high-stakes domain — structured pipelines, specialist routing, auditing, and honest disclaimers over black-box answers.

Key Outcomes

  • Seven-stage diagnostic pipeline: triage, parallel diagnosis and research, specialist consult, treatment, safety, and empathy agents
  • Five medical specialist sub-agents (cardiology, neurology, dermatology, gastroenterology, psychiatry) consulting over an async message bus
  • Multi-provider LLM layer — Claude primary, with OpenAI and Gemini support and Whisper voice transcription
  • Multimodal intake: chat, voice, photo upload with annotation, and an interactive body map
  • Server-sent events stream live pipeline progress; a self-learning quality auditor reviews every diagnosis