CareGraph

AI-powered senior care intelligence — Neo4j, RocketRide AI, GMI Cloud, Bland AI, and CrewAI.

The problem

Clinical care is infrequent; symptoms and medications change daily. Elders risk missed side effects and late escalation; family caregivers lack one trustworthy, connected picture. CareGraph adds daily AI check-ins, a Neo4j graph of meds, symptoms, and relationships, and alerts so issues surface before an emergency.

Demo: https://caregraph.onrender.com (hosted snapshot for UI exploration — use Quick Start for a full local stack).

Neo4j and RocketRide AI

Neo4j stores the care graph (seniors, meds, symptoms, check-ins, doctors, alerts). The app queries it with Cypher to connect events quickly, detect risks like side effects/interactions, and power dashboard insights and recommendations.

RocketRide AI runs reusable .pipe workflows for transcript analysis, drug explanations, care plans, and condition suggestions. This keeps AI reasoning consistent and modular, with app/services/rocketride.py falling back to GMI Cloud when RocketRide is unavailable.

Quick Start

Prerequisites

Python 3.12+
uv (package manager)
Neo4j Aura account (or local Docker) — console.neo4j.io
GMI Cloud API key — console.gmicloud.ai
Bland AI API key (optional, for voice calls) — bland.ai

1. Clone and install

git clone https://github.com/SankarSubbayya/CareGraph.git
cd CareGraph
uv sync

2. Configure environment

Copy .env.example or create .env:

# Neo4j (Aura or local)
NEO4J_URI=neo4j+s://your-instance.databases.neo4j.io
NEO4J_USER=neo4j
NEO4J_PASSWORD=your-password

# GMI Cloud (required for AI features)
GMI_BASE_URL=https://api.gmi-serving.com/v1
GMI_API_KEY=your-gmi-api-key
GMI_MODEL=Qwen/Qwen3-235B-A22B-Instruct-2507-FP8

# Bland AI (optional, for voice calls)
BLAND_API_KEY=your-bland-api-key

# RocketRide AI (optional, for pipeline orchestration)
ROCKETRIDE_URI=http://localhost:5565
ROCKETRIDE_APIKEY=

# App
BASE_URL=http://localhost:8000
SKIP_AUTH=true

Using local Neo4j instead of Aura:

docker run -d --name neo4j -p 7474:7474 -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/password neo4j:5

Then set NEO4J_URI=bolt://localhost:7687 and NEO4J_PASSWORD=password.

3. Seed demo data and start

# Start the server
uv run python main.py

# In another terminal — seed demo data
uv run python scripts/seed_data.py     # 4 seniors, medications, check-ins, alerts
uv run python scripts/seed_doctors.py  # 159 doctors, 38 clinics

4. Open in browser

URL	Page
http://localhost:8000	Landing page
http://localhost:8000/dashboard	Full dashboard

5. Run tests

uv run python -m pytest tests/ -v

What CareGraph Does

1. Bland AI calls the senior every morning
2. Voice agent asks about mood, medications, symptoms, doctor needs
3. Transcript is analyzed → symptoms extracted → stored in Neo4j graph
4. Graph detects drug interactions, side effect matches, condition suggestions
5. GMI Cloud (Qwen3-235B) generates care plans from graph data
6. Alerts notify family members based on severity

Example: Dorothy reports dizziness

Dorothy TAKES Lisinopril
Dorothy REPORTED dizziness
Lisinopril HAS_SIDE_EFFECT dizziness
→ Neo4j connects the dots automatically
→ Qwen3-235B explains: "Dizziness may be a side effect of Lisinopril. Discuss with doctor."
→ Family gets notified

Architecture

flowchart TB
    subgraph Users["Users"]
        Family["Family\n(Dashboard)"]
        SeniorPhone["Senior\n(Phone)"]
    end

    subgraph BlandAI["Bland AI (Voice Agent)"]
        VoiceCall["Automated\nPhone Call"]
        Webhook["Webhook\n(Transcript)"]
    end

    subgraph CrewAI["CrewAI (Multi-Agent Orchestration)"]
        Agent1["Check-in Agent\n(Bland AI calls)"]
        Agent2["Analysis Agent\n(NLP extraction)"]
        Agent3["Graph Agent\n(Neo4j queries)"]
        Agent4["Recommendation Agent\n(AI care plans)"]
        Agent5["Alert Agent\n(Safety monitor)"]

        Agent1 -->|transcript| Agent2
        Agent2 -->|symptoms, mood| Agent3
        Agent3 -->|graph insights| Agent4
        Agent4 -->|care plan| Agent5
    end

    subgraph API["FastAPI Backend"]
        CRUD["Senior CRUD"]
        CheckinAPI["Check-in\nProcessing"]
        GraphAPI["Graph\nIntelligence"]
        AlertAPI["Alert Engine"]
        VoiceAPI["Voice\nEndpoints"]
        CrewAPI["Crew\nEndpoints"]
    end

    subgraph Neo4j["Neo4j Aura (Graph Database)"]
        Senior["(:Senior)"]
        Med["(:Medication)"]
        Sym["(:Symptom)"]
        Cond["(:Condition)"]
        CI["(:CheckIn)"]
        Alert["(:Alert)"]
        Fam["(:FamilyMember)"]
        Svc["(:Service)"]

        Senior -->|"TAKES"| Med
        Senior -->|"REPORTED"| Sym
        Senior -->|"CHECKED_IN"| CI
        Senior -->|"HAS_CONTACT"| Fam
        Senior -->|"NEEDS"| Svc
        CI -->|"DETECTED"| Sym
        CI -->|"TRIGGERED"| Alert
        Med -->|"INTERACTS_WITH"| Med
        Med -->|"SIDE_EFFECT"| Sym
        Sym -->|"SUGGESTS"| Cond
    end

    subgraph Inference["LLM Inference"]
        RocketRide["RocketRide AI\n(.pipe pipelines)"]
        GMI["GMI Cloud\n(Qwen3-235B)"]
        RocketRide -.->|fallback| GMI
    end

    Family --> API
    SeniorPhone <--> BlandAI
    BlandAI -->|webhook| API
    API --> CrewAI
    CrewAI --> Neo4j
    CrewAI --> Inference
    CrewAI --> BlandAI
    Agent1 -->|"initiate call"| VoiceCall
    Webhook -->|"transcript"| CheckinAPI
    Agent3 -->|"Cypher queries"| Neo4j
    Agent4 -->|"prompts"| Inference
    Agent5 -->|"alerts"| AlertAPI
    GraphAPI -->|"Query"| Neo4j
    GraphAPI -->|"Reason"| Inference

Data Flow

sequenceDiagram
    participant F as Family Dashboard
    participant A as FastAPI
    participant C as CrewAI Crew
    participant B as Bland AI
    participant S as Senior (Phone)
    participant N as Neo4j Aura
    participant L as GMI Cloud LLM

    F->>A: POST /api/crew/checkin/{phone}
    A->>C: Start Full Check-in Crew
    C->>N: Look up senior profile
    N-->>C: Name, medications, contacts
    C->>B: Initiate voice call
    B->>S: Automated phone call
    S-->>B: Conversation (mood, meds, symptoms)
    B-->>A: Webhook: transcript + recording
    A->>C: Analysis Agent processes transcript
    C->>N: Store check-in + symptoms in graph
    C->>N: Query drug interactions & side effects
    N-->>C: Graph insights (interactions, matches)
    C->>L: Generate care recommendations
    L-->>C: Personalized care plan
    C->>N: Evaluate & store alerts
    C-->>A: Complete crew output
    A-->>F: Results + alerts + care plan

Tech Stack

Layer	Technology	Role
Graph Database	Neo4j Aura	Knowledge graph — 10 node types, 159 doctors, 38 clinics, 1000+ relationships
Voice Agent	Bland AI	Automated phone calls to seniors with doctor recommendations
AI Pipelines	RocketRide AI	Visual pipeline orchestration (.pipe files)
LLM Inference	GMI Cloud (Qwen3-235B)	235B parameter model for care plans, drug explanations
Agent Orchestration	CrewAI	5 specialized agents with 11 custom tools
Backend	FastAPI	Python REST API — 33 endpoints
Frontend	HTML/JS/vis.js	Interactive dashboard with graph visualization

Graph Model

(:Senior)-[:TAKES]->(:Medication)
(:Senior)-[:REPORTED]->(:Symptom)
(:Senior)-[:CHECKED_IN]->(:CheckIn)-[:DETECTED]->(:Symptom)
(:Senior)-[:HAS_CONTACT]->(:FamilyMember)
(:Senior)-[:NEEDS]->(:Service)
(:Medication)-[:INTERACTS_WITH]->(:Medication)
(:Medication)-[:SIDE_EFFECT]->(:Symptom)
(:Symptom)-[:SUGGESTS]->(:Condition)
(:Condition)<-[:CAN_TREAT]-(:Doctor)
(:Doctor)-[:PRACTICES_AT]->(:Clinic)
(:CheckIn)-[:TRIGGERED]->(:Alert)

Dashboard Pages

Page	Features
Home	Landing page — problem statement, solution flow, live Neo4j stats
Seniors	List seniors, wellness scores, family contacts, action buttons
Graph View	Interactive vis.js graph — Care Network + Doctors Network views
Graph Reasoning	Animated step-by-step walkthrough of Neo4j reasoning chain
AI Insights	Drug interactions, side effects, condition suggestions, doctor recommendations, cross-senior search
Voice Calls	Initiate Bland AI calls, voice selection, call history, save to graph
CrewAI Agents	Visual 5-agent pipeline, run full check-in / analyze / insights
Alerts	Severity-coded alerts with family notification targets
Simulate	Enter transcript, see analysis + alerts + family notifications

Key Demo Scenarios

Scenario	What Neo4j Does	What AI Does
Margaret takes Metformin + Lisinopril	Detects INTERACTS_WITH relationship	Qwen3-235B explains the interaction risk
Dorothy reports dizziness	Matches symptom to Lisinopril SIDE_EFFECT	Suggests talking to doctor
3 seniors report similar symptoms	Finds shared symptom paths in graph	Identifies potential cause
Senior needs a doctor	Traverses Symptom → Condition → Doctor → Clinic	Recommends specific doctors

API Endpoints (33 total)

Seniors

POST /api/seniors — Add senior
GET /api/seniors — List all
GET /api/seniors/{phone} — Get one
DELETE /api/seniors/{phone} — Remove

Check-ins

POST /api/checkins/simulate/{phone} — Simulate with transcript
GET /api/checkins/{phone} — History
GET /api/checkins/latest/all — Latest per senior

Graph Intelligence

GET /api/graph/stats — Live graph statistics
GET /api/graph/care-network/{phone} — Care network visualization
GET /api/graph/doctors-network/{phone} — Doctors network visualization
GET /api/graph/drug-interactions/{phone} — Drug interactions + AI explanation
GET /api/graph/side-effects/{phone} — Side effect matches
GET /api/graph/similar-symptoms/{phone} — Cross-senior symptom patterns
GET /api/graph/condition-suggestions/{phone} — AI condition suggestions
GET /api/graph/care-recommendation/{phone} — AI care plan
GET /api/graph/doctors — Search doctors by specialty/city
GET /api/graph/doctors/for-senior/{phone} — Recommended doctors
GET /api/graph/seniors-by-symptom/{symptom} — Find by symptom
GET /api/graph/seniors-by-medication/{med} — Find by medication

Voice Agent (Bland AI)

POST /api/voice/call/{phone} — Call a senior
POST /api/voice/call-all — Call all seniors
GET /api/voice/call/{call_id} — Call details + transcript
POST /api/voice/call/{call_id}/analyze — Post-call AI analysis
POST /api/voice/call/{call_id}/stop — Stop call
POST /api/voice/process/{call_id} — Save call transcript to graph
GET /api/voice/calls — Recent calls
POST /api/voice/webhook — Bland AI callback

CrewAI Agents

POST /api/crew/checkin/{phone} — Full 5-agent pipeline
POST /api/crew/analyze/{phone} — Analysis pipeline (4 agents)
POST /api/crew/insights/{phone} — Graph insights (2 agents)

Alerts

GET /api/alerts — Active alerts
PUT /api/alerts/{id}/acknowledge — Acknowledge

RocketRide AI Pipelines

4 visual pipelines in pipelines/ directory:

Pipeline	Purpose
`checkin_analysis.pipe`	Transcript → symptoms, mood, urgency
`drug_interaction.pipe`	Drug pair → plain-language explanation
`care_recommendation.pipe`	Graph data → personalized care plan
`condition_suggestion.pipe`	Symptom cluster → possible conditions

Each follows: Webhook → Prompt → Gemini LLM → Response

Setup: Install RocketRide VS Code extension → Open .pipe file → Configure Gemini key → Click play

Inference chain: RocketRide pipeline → GMI Cloud (Qwen3-235B) fallback → empty

CrewAI Agents

5 agents collaborate on every check-in:

Check-in Agent → Analysis Agent → Graph Agent → Recommendation Agent → Alert Agent
  (Bland AI)      (NLP extract)    (Neo4j)       (Qwen3-235B)          (Alerts)

Agent	Tools
Check-in Agent	Bland AI voice calls, senior lookup
Analysis Agent	NLP transcript analyzer, Neo4j store
Graph Agent	Drug interactions, side effects, similar symptoms, care network
Recommendation Agent	GMI Cloud LLM for explanations and care plans
Alert Agent	Severity evaluation, family notification

Project Structure

CareGraph/
├── main.py                        # FastAPI app entry point
├── .env                           # Configuration (gitignored)
├── pipelines/                     # RocketRide AI pipelines
│   ├── checkin_analysis.pipe
│   ├── drug_interaction.pipe
│   ├── care_recommendation.pipe
│   └── condition_suggestion.pipe
├── app/
│   ├── config.py                  # Pydantic settings
│   ├── graph_db.py                # Neo4j Cypher queries (467 lines)
│   ├── crew/                      # CrewAI multi-agent system
│   │   ├── agents.py              # 5 agent definitions
│   │   ├── tasks.py               # Task definitions
│   │   ├── tools.py               # 11 custom tools
│   │   └── care_crew.py           # 3 crew pipelines
│   ├── models/
│   │   └── senior.py              # Pydantic models
│   ├── routers/
│   │   ├── seniors.py             # Senior CRUD
│   │   ├── checkins.py            # Check-in processing
│   │   ├── alerts.py              # Alert management
│   │   ├── graph.py               # Graph intelligence + AI
│   │   ├── voice.py               # Bland AI voice endpoints
│   │   └── crew.py                # CrewAI endpoints
│   └── services/
│       ├── bland_voice.py         # Bland AI client + doctor lookup
│       ├── rocketride.py          # RocketRide + GMI Cloud fallback
│       ├── gmi_inference.py       # GMI Cloud API client
│       ├── call_analyzer.py       # Local NLP analysis
│       └── alert_engine.py        # Alert rules + family notifications
├── frontend/
│   ├── landing.html               # Landing page
│   ├── index.html                 # Dashboard
│   ├── app.js                     # Frontend logic
│   └── style.css                  # Styles
├── scripts/
│   ├── seed_data.py               # Demo seniors + medical knowledge
│   └── seed_doctors.py            # 159 doctors + 38 clinics
├── tests/                         # 60 tests (unit + integration)
├── data/                          # EHR sample data
└── presentation/
    └── DEMO_SCRIPT.md             # 10-slide demo script

Tests

60 tests — all passing:

uv run python -m pytest tests/ -v

Test File	Count	What
test_models.py	3	Pydantic models
test_call_analyzer.py	19	NLP: mood, meds, symptoms, services
test_alert_engine.py	13	Alert rules, severity, dedup, source keys
test_config.py	3	Settings defaults, overrides, Aura alias normalization
test_integration.py	22	Neo4j queries, API endpoints, full pipelines

Medical Disclaimer

All AI-generated content includes disclaimers:

System prompt forces: "This is AI-generated guidance. Always consult your doctor for medical decisions."
Drug interactions: "Consult your doctor before making any medication changes."
Condition suggestions: "Consult your doctor for proper diagnosis and treatment."
Bland AI voice agent: "Your doctor would know best about your specific situation."
Frontend: Yellow disclaimer banner on all AI results pages

Security

.env is gitignored — never committed
.pipe files use ${ROCKETRIDE_GEMINI_APIKEY} env var reference — no hardcoded keys
.env.example provided with <YourKeyHere> placeholders for deployment
Optional demo auth via DEMO_USERNAME / DEMO_PASSWORD (browser basic auth)
Admin endpoints protected by ADMIN_API_TOKEN
See PUBLIC_DEMO.md for deployment guide

Deployment

Render (one-click)

render.yaml blueprint included. See PUBLIC_DEMO.md for details.

Environment Variables

See .env.example for all required and optional variables.

Open Source Contribution

We contributed a Bland AI tool node to the RocketRide project:

PR: rocketride-org/rocketride-server#521
Adds make_call, get_call, analyze_call tools for RocketRide agents

Changelog

Commit	Change
Medical disclaimers	All AI responses include "consult your doctor" warnings; frontend yellow banners
Config tests fix	Clear env vars before assertions for AliasChoices compatibility
Pipe env vars	`.pipe` files use `${ROCKETRIDE_GEMINI_APIKEY}` — no hardcoded API keys
Production merge	Security middleware, Render deploy, alert dedup, Aura NEO4J_USERNAME alias
AI insights enhanced	Condition suggestions, recommended doctors, cross-senior search by symptom/medication
Home link	Dashboard logo + sidebar link back to landing page
Demo mode fix	Non-blocking toast in bottom-right corner
Graph reasoning	Animated 6-step walkthrough of Neo4j reasoning chain with vis.js graph
Family notifications	Alerts notify emergency contacts by severity (critical → all, high → primary)
Live stats	Landing page shows real-time Neo4j counts
Demo mode	One-click 8-step automated dashboard walkthrough
Docs + Qwen3-235B	Updated all docs for GMI Cloud Qwen3-235B model
Doctors graph	159 doctors + 38 clinics in Neo4j; interactive doctors network visualization
60 tests	Unit + integration tests all passing
CrewAI	5 agents, 11 tools, 3 crew pipelines
Bland AI voice	Automated check-in calls with doctor recommendations from Neo4j
RocketRide pipelines	4 .pipe files for AI reasoning
GMI Cloud	Qwen3-235B inference for care plans, drug explanations
Initial	Neo4j graph model, FastAPI backend, dashboard frontend

Name		Name	Last commit message	Last commit date
Latest commit History 40 Commits
app		app
data		data
frontend		frontend
outputs		outputs
pipelines		pipelines
presentation		presentation
scripts		scripts
tests		tests
.env.example		.env.example
.gitignore		.gitignore
.python-version		.python-version
NEO4J_ARCHITECTURE.md		NEO4J_ARCHITECTURE.md
PUBLIC_DEMO.md		PUBLIC_DEMO.md
README.md		README.md
REQUIREMENTS.md		REQUIREMENTS.md
main.py		main.py
pyproject.toml		pyproject.toml
render.yaml		render.yaml
uv.lock		uv.lock

Folders and files

Latest commit

History

Repository files navigation

CareGraph

The problem

Neo4j and RocketRide AI

Quick Start

Prerequisites

1. Clone and install

2. Configure environment

3. Seed demo data and start

4. Open in browser

5. Run tests

What CareGraph Does

Example: Dorothy reports dizziness

Architecture

Data Flow

Tech Stack

Graph Model

Dashboard Pages

Key Demo Scenarios

API Endpoints (33 total)

Seniors

Check-ins

Graph Intelligence

Voice Agent (Bland AI)

CrewAI Agents

Alerts

RocketRide AI Pipelines

CrewAI Agents

Project Structure

Tests

Medical Disclaimer

Security

Deployment

Render (one-click)

Environment Variables

Open Source Contribution

Changelog

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages