Agentic Pharmacy AI (36h Hackathon)

This repository contains an autonomous, intelligent multi-agent pharmacy system built entirely on the 3-Layer Agent Architecture (Directives, Orchestrator, Execution).

Features

1. Agentic Chain of Thought (Observability): Integrated with LangGraph and LangSmith. Provides full trace logs proving the Supervisor checked Inventory AND Prescriptions before fulfilling orders.
2. Conversational Ordering (STT/TTS): A React frontend that uses Web Speech API to capture audio, passes it to the FastAPI LangGraph backend, and speaks the response.
3. Safety Guardrails: Strict Python deterministic tools that check MongoDB stock limits and Mock prescription limits.
4. Proactive Refill Intelligence: A CRON background task that mathematically predicts when patients will run out of chronic medications based on sandbox purchase history.
5. No-CMS Sandbox: Entirely custom FastAPI + MongoDB backend, replacing SQLite for superior async performance per requirements.

Stack

• AI/Orchestration: LangGraph, LangChain, gpt-4o-mini (or local Ollama fallback)
• Backend: Python, FastAPI, Motor (Async MongoDB), APScheduler (for refill crons).
• Frontend: React (Vite), TailwindCSS, lucide-react.
• Database: MongoDB
• Observability: LangSmith

How to Run

1. Database & Environment

1. Ensure MongoDB is running locally on port 27017.
2. Open .env and populate your OPENAI_API_KEY (or swap to local ChatOllama in agent_supervisor.py).
3. (Optional) Populate LANGCHAIN_API_KEY in .env to view trace chains live in the browser.

2. Backend (FastAPI / LangGraph)

1. Navigate to the backend/ directory (or use the root venv).
2. Run data load (assuming you place mock excels in /data): python execution/data_loader.py
3. Boot the server: uvicorn backend.main:app --reload --port 8000

3. Frontend (React / Vite)

1. Open a second terminal to /frontend
2. npm install
3. npm run dev
4. Access the UI at http://localhost:5173.
   • Click "Test Chatbot" to place voice orders.
   • Click "Admin View" to see live inventory depletion and CRON-generated refill alerts.

3-Layer Architecture Example

When you say "I need 20 pills of Paracetamol":

1. Orchestrator (LangGraph) reads the NLP and checks directives/agent_supervisor_sop.md.
2. It invokes the LangChain Tool check_inventory which fires an HTTP request to our determinist API.
3. It invokes check_prescription based on directives/prescription_validation.md.
4. If BOTH pass, it invokes execute_order, which writes to MongoDB.
5. The CRON job in predict_refills.py continuously scans the new DB state to forecast future shortages.

Give a Star 🌟 to this repository!!