Healthcare IT Decision-Making (Gradio + crewAI)

Repository for a multi-agent healthcare IT decision-making app.

• UI: Gradio with a single text input + report output
• Orchestration: crewAI with 3 specialized agents
• Search: serper.dev (Google results API) via SerperDevTool
• Local RAG: simple CSV (company_data.csv) to ground budget/risk context
• Output: Executive-ready Markdown report rendered in the app

Built as a Cursor project.

---

📸 Screenshots

---

🧰 Tech Stack

• Python 3.11.9+
• crewai, crewai-tools (SerperDevTool), gradio, pandas, python-dotenv

---

🚀 Getting Started

1. Clone this repo and enter the folder:

   git clone <your-repo-url>.git
   cd healthcare-it-decision-app

2. Install dependencies:

   pip install -r requirements.txt

3. Set API keys (use a .env file or environment variables):

   • OPENAI_API_KEY : for LLM access (used by crewAI)
   • SERPER_API_KEY : for serper.dev web search
   • Optional:
     • CREWAI_MODEL (default: gpt-4o-mini)
     • CREW_MAX_TOKENS (default: 2000)
     • LOCAL_DATA_PATH (default: company_data.csv)

   Example .env:

   OPENAI_API_KEY=sk-your-openai-key
   SERPER_API_KEY=your-serper-key
   CREWAI_MODEL=gpt-4o-mini
   CREW_MAX_TOKENS=2000

4. Run the app:

   python app.py

5. Open the local Gradio URL printed in the console.

---

🧪 Try These Example Prompts

• Evaluate the pros and cons of migrating our EHR to the cloud (AWS vs Azure).
• Assess the 3-year TCO for building a new patient-facing mobile app.
• Should we replace our on-prem PACS with a cloud archive?

---

🧩 Local RAG (CSV) — Customization

• Edit company_data.csv to reflect your environment (systems, costs, risk).
• The LocalDataTool searches this file by keyword and returns a compact JSON with the top matches + cost/risk rollups.
• To wire a real database or a vector store, replace the tool's implementation in app.py:
  • Keep the function signature the same.
  • Use your DB driver / retrieval client inside.
  • Return JSON so agents can reason over it.

---

🧱 Repository Layout

healthcare-it-decision-app/
├─ app.py                # single-file app with crewAI + Gradio
├─ requirements.txt
├─ company_data.csv      # sample data source used by LocalDataTool
├─ README.md
└─ images/
   └─ mock_screenshot.png

---

🔧 Configuration Notes

• No hard-coded secrets. All API keys are read from env or .env (via python-dotenv).
• Output is Markdown rendered directly in the Gradio UI.
• crewAI is configured with a sequential process so agents build on each other's outputs.

---

🧠 Agents Overview

• ResearchAgent — Market/standards research with serper.dev
• FinancialAnalystAgent — TCO/ROI using web + CSV
• RiskAssessorAgent — Implementation/security risks using web + CSV

The final synthesis task compiles all sections into an executive-ready report with headings:

1. Summary Recommendation
2. Budget/Cost Analysis
3. Potential Downtime & Business Impact
4. Implementation Risks
5. Technological Fit & Future-Proofing
6. Sources

---

🧱 Limitations & Next Steps

• This repo is a demo. Cost figures are estimates; always validate with vendors.
• For production, consider:
  • Audit logging & PHI redaction
  • Model/Tool observability (e.g., Langfuse)
  • Persistent vector DB for richer RAG
  • Caching for deterministic demos
  • Fine-grained prompts & evals

---

📣 Attribution

Developed by Jeremey King in Cursor.

---

🐳 Docker (optional)

Build and run with Docker:

docker build -t healthcare-it-decision-app .
docker run --rm -p 7860:7860 --env-file .env healthcare-it-decision-app

🛠 Make targets

make install
make run
make docker-build
make docker-run

🔐 Environment file

Start from .env.example:

cp .env.example .env
# fill in keys

---

🤖 Using Different Models (Gemini, Llama, etc.)

This app uses crewAI’s LiteLLM integration. Set CREWAI_MODEL to a provider-prefixed model string and provide the matching API key env var.

Examples

• OpenAI (default):
  CREWAI_MODEL=openai/gpt-4o-mini | Key: OPENAI_API_KEY
• Google Gemini:
  CREWAI_MODEL=gemini/gemini-1.5-pro | Key: GEMINI_API_KEY
• Meta Llama via Groq:
  CREWAI_MODEL=groq/llama-3.1-70b-versatile | Key: GROQ_API_KEY
• Meta Llama via Together:
  CREWAI_MODEL=together_ai/meta-llama/Llama-3-70b-instruct | Key: TOGETHER_API_KEY
• Meta Llama via OpenRouter:
  CREWAI_MODEL=openrouter/meta-llama/llama-3.1-70b-instruct | Key: OPENROUTER_API_KEY

Tip: If your runs fail with “LLM Provider NOT provided,” you’re likely missing the provider prefix (e.g., openai/, gemini/, groq/, together_ai/, openrouter/).
Use Python 3.11+ for best compatibility.

Minimal .env example for Gemini:

GEMINI_API_KEY=your-gemini-key
CREWAI_MODEL=gemini/gemini-1.5-pro
SERPER_API_KEY=your-serper-key

Minimal .env example for Groq + Llama:

GROQ_API_KEY=your-groq-key
CREWAI_MODEL=groq/llama-3.1-70b-versatile
SERPER_API_KEY=your-serper-key

---

🔎 serper.dev account (web search)

This app uses serper.dev for real-time web search. Create a free account at serper.dev and generate an API key, then set:

SERPER_API_KEY=your-serper-key

As of now, the first 2,500 queries are free (no credit card required). Check their site for current limits and pricing.