🩺 BioGraph - Medical Biomarker Analysis Assistant

A specialized AI-powered medical assistant built on Cloudflare's Agent platform, designed to analyze biomarkers and provide detailed medical insights. This project provides an interactive chat interface for medical professionals to get thorough analysis of biomarker data with benchmarking and recommendations.

Features

• 🔬 Comprehensive biomarker analysis
• 📊 Benchmark comparison
• 💬 Interactive medical consultation interface
• 🧠 Powered by xAI Grok 4.6 on Amazon Bedrock
• 🎯 Detailed action plans and next steps
• 🌓 Dark/Light theme support
• ⚡️ Real-time streaming responses
• 🔄 Conversation history tracking

A lab-panel question renders as a table (marker, value, unit, range, flag) plus next-step prose and follow-up chips. Questions with no lab values stay plain chat. While a reply is in flight the UI shows a waiting row; if the model call fails it shows an error with retry, not a blank thread.

Configuration

Variable	Required	Default	Notes
OPENAI_API_KEY	yes	—	Bedrock mantle API key. Not an xAI key, and not a bedrock-runtime key.
OPENAI_BASE_URL	no	https://bedrock-mantle.us-west-2.api.aws/openai/v1
OPENAI_MODEL_ID	no	xai.grok-4.6	Set to xai.grok-4.3 to switch models.
OPENAI_REASONING_EFFORT	no	none	none / low / medium / high. Anything else falls back to none.

Local config goes in .dev.vars (gitignored). Production: pnpm wrangler secret put OPENAI_API_KEY.

reasoning_effort is the dominant cost lever on Grok 4.6 — reasoning tokens bill at the output rate and are never returned to you. Measured on a 7-marker panel question: low (the model default) cost $0.01107 per response, none cost $0.00341 with no loss of accuracy. Shortening the system prompt does not save money; it made the model reason longer and cost 56% more.

Requests are rate limited to 20 per minute, both per client IP at the edge and per conversation inside the Durable Object. There is no authentication — anyone with the URL can use the app, within that limit.

Note the rate-limit binding is not wired by vite dev (@cloudflare/vite-plugin 0.1.x), so limits are inactive there and the app degrades to allowing every request. wrangler dev and production both bind it.

Testing

pnpm run check   # prettier + biome + tsc
pnpm test        # unit tests
pnpm run start   # dev server
pnpm run smoke   # end-to-end: drives the real agent over its WebSocket

Use Cases

1. Biomarker Analysis

   • Compare lab results against established benchmarks
   • Identify out-of-range values
   • Get detailed explanations of implications

2. Medical Recommendations

   • Receive evidence-based next steps
   • Get detailed action plans
   • Access follow-up recommendations

3. Medical Professional Support

   • Quick access to biomarker interpretations
   • Evidence-based decision support
   • Efficient patient data analysis

License

MIT