CareCompanion - Agentic Healthcare AI Assitant

CareCompanion AI is an autonomous agentic system that acts as a personal health companion for patients managing Type 2 Diabetes. Unlike a regular chatbot, it remembers patients across sessions, calls real healthcare APIs to fetch live data, and proactively surfaces FDA drug alerts without being asked. The agent reasons across three domains simultaneously - medications, nutrition, and cost — and generates personalized responses grounded in the patient's actual medical profile.

Key Features

• Persistent memory - remembers patient history across sessions using ChromaDB vector store + SQLite
• 5 real healthcare APIs - OpenFDA, RxNorm, USDA FoodData, PubMed, GoodRx pricing
• Autonomous tool selection - agent decides which APIs to call based on patient query
• Proactive FDA monitoring - surfaces drug safety alerts without being asked
• Cost optimization - identifies generic medication alternatives and annual savings
• Nutrition assessment - flags foods by glycemic impact for diabetes management
• Multi-patient support - 70 simulated patient profiles with realistic clinical data

Architecture

User Message
     │
     ▼
┌─────────────────┐
│  Node 1         │  Load patient profile (SQLite)
│  load_context   │  Retrieve relevant past conversations (ChromaDB)
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  Node 2         │  LLM decides which tools are needed
│  decide_tools   │  based on patient profile + message
└────────┬────────┘
         │
         ▼
┌─────────────────┐     ┌──────────────────────────────────────┐
│  Node 3         │────▶│  OpenFDA  · RxNorm  · USDA           │
│  run_tools      │     │  PubMed   · GoodRx pricing           │
└────────┬────────┘     └──────────────────────────────────────┘
         │
         ▼
┌─────────────────┐
│  Node 4         │  Generates personalized response
│  generate       │  Saves to memory · Writes agent note
└─────────────────┘

Memory Architecture:

SQLite — structured patient facts (medications, glucose, BMI, budget, insurance) ChromaDB — conversation history as vectors for semantic search across sessions

Tech Stack

• LLM: Groq (Llama 3.3 70B)
• Framework: LangGraph
• Memory & Storage: ChromaDB + SQLite
• APIs: OpenFDA, RxNorm, USDA FoodData Central, PubMed
• UI: Gradio

Evaluation Metrics

• Tool Call Success Rate: 81.8%
• Memory Retrieval Accuracy: 66.7%
• Avg Annual Cost Savings: $2,294 per patient
• Nutrition Assessment Accuracy: 70%
• FDA Alert Coverage: 100%

Project Structure

CareCompanion/
│
├── codes/
│   ├── data.ipynb       
│   ├── memory_arch.ipynb  
│   ├── api.ipynb       
│   ├── agent.ipynb             
│   ├── ui.ipynb            
│   └── evaluation.ipynb           
│
├── data/
│   ├── raw/                          # Pima Diabetes dataset (768 patients)
│   └── processed/
│       ├── diabetes_clean.csv        
│       └── distributions.png         
│
├── memory/
│   ├── sqlite/carecompanion.db       # Patient profiles + structured facts
│   └── chroma_db/                    # Conversation vector store
│
├── evaluation/
│   ├── test_patients.json            
│   ├── test_patients_expanded.json   
│   ├── metrics_report.json           
│   └── metrics_dashboard.png        
│
└── requirements.txt

Dataset

• Pima Indians Diabetes Database — 768 real patient records (Kaggle/UCI)
• 70 simulated test patients — derived from Pima dataset with realistic medication regimens, insurance types, budgets, and comorbidities
• USDA FoodData Central — 300k+ foods with full nutritional profiles

Quickstart

1. Clone the repo

git clone https://github.com/maithili74/CareCompanion
cd CareCompanion

3. Install dependencies

pip install -r requirements.txt

5. Get free API keys

Groq: https://console.groq.com (instant, no card) USDA: https://fdc.nal.usda.gov/api-guide.html (instant) OpenFDA: no key needed PubMed: no key needed

4. Set your keys

os.environ["GROQ_API_KEY"] = "your_key_here"
os.environ["USDA_API_KEY"] = "your_key_here"

5. Run notebooks in order

data -> memory_arch -> api -> agent -> ui

6. Launch the UI Run all cells in ui.ipynb, a local + public URL will appear automatically.