docs/plans/mvp_checklist.md

• Initialize new Hugging Face Space with Gradio SDK 5.x

  • Add mcp-server-track tag in README.md

• Write Python function symptom_to_diagnosis(symptom_text)

  • Use OpenAI or Anthropic API to generate JSON
  • Format prompt to request JSON output
  • Parse model response into Python dict
  • Handle JSON formatting quirks (trim extra text, use json.loads)
  • Implement fallback rule-based mapping for demo cases

• Test symptom_to_diagnosis function

  • Input common symptom examples and combinations
  • Verify relevance and correctness of ICD codes and diagnoses
  • Tweak prompt to improve specificity and JSON validity

• Define confidence score methodology

  • Decide whether to use model’s self-reported scores or rank order as proxy
  • Document how confidence is calculated and interpreted

• Integrate function into Gradio Blocks interface

  • Use gr.Interface or gr.ChatInterface to accept symptom text input and display JSON output
  • Configure Gradio app metadata to expose MCP endpoint

• Build demonstration client or script (optional)

  • Create minimal client using gradio.Client or requests to call Space’s prediction API
  • Alternatively, build a second Gradio Space as a simple chatbot that calls the MCP tool
  • Prepare screen recording showing AI agent (e.g., Claude Desktop) calling the MCP endpoint with example query

• Update README.md documentation

  • Describe tool functionality and usage examples
  • Include mcp-server-track tag, link to video or client demo
  • List technologies used (e.g., “OpenAI GPT-4 API for symptom→ICD mapping”)

• Configure OpenAI/Anthropic API usage

  • Use cheaper models (e.g., GPT-3.5) during development
  • Reserve GPT-4 or Claude-2 for final demo queries to conserve credits

• Evaluate Hugging Face / Mistral credits for alternative inference

  • Identify open ICD-10 prediction models on HF Inference API (e.g., AkshatSurolia/ICD-10-Code-Prediction)
  • Consider running open-source models on Mistral if time allows

• Plan Modal Labs usage for cloud compute (optional)

  • Pre-compute ICD-10 embeddings in Modal job if semantic search is added
  • Host backend microservice or Gradio app on Modal if HF Space resources are insufficient

• Reserve Nebius or Hyperbolic Labs credits for GPU-intensive tasks (if needed)

  • Spin up GPU instance to host or fine-tune open-source model only if HF Space times out

• Consider LlamaIndex integration for retrieval-augmented generation (bonus)

  • Load ICD-10 dataset into LlamaIndex and test semantic search for candidate codes
  • Implement minimal index of common diagnoses for demo if time permits

• Record and document final demo

  • Capture symptom input, MCP tool invocation, and JSON output in a short video
  • Host video link in README.md