❤️‍🩹 CareMate - Let's Take Care of Caregivers!

Loom Demo Video || Slide Deck || GitHub Repo\

Inspiration

63 million Americans provide unpaid caregiving—that's 1 in 5 adults. 61% are women, and nearly half are "sandwich generation" caregivers juggling children, aging parents, and full-time careers. The statistics are staggering: 47% struggle with burnout, 43% are chronically sleep-deprived, and caregivers lose up to 90% of their retirement savings compared to non-caregivers.

The invisible crisis? No platform detects burnout before it becomes a health emergency.

Existing solutions fall short:

Task apps (CareZone, Caring Village) track medication reminders but don't detect burnout
Wellness apps (Calm, Headspace) offer generic meditation tips without caregiver context
Wearable apps (Garmin, Fitbit) collect physiological data but provide no actionable insights for caregivers
Healthcare platforms focus on care recipients, not the caregivers themselves

CareMate is the first platform to combine objective biomarkers from wearables with AI-driven interventions to prevent caregiver burnout before crisis. We transform calendar density, task overload, heart rate variability, sleep patterns, and stress markers into real-time risk scoring and hyper-specific, low-friction interventions.

What it does

CareMate is a multimodal AI platform that detects caregiver burnout in real-time by synthesizing multiple data streams into a unified 0-100 risk score, then generates actionable interventions tailored to each caregiver's specific situation.

Core Features

🎯 Real-Time Burnout Detection

4-component risk scoring algorithm combining wearable biomarkers (30%), calendar/task load (30%), semantic analysis (25%), and self-reports (15%)
Live score updates when new data arrives (health uploads, calendar changes, mood check-ins)
Color-coded risk levels: Managing (0-25), Moderate (26-50), High (51-70), Critical (71-85), Crisis (86-100)
Historical trend tracking with daily burnout, sleep, and HRV charts

📊 Wearable Health Integration

JSON health data upload capturing heart rate variability (HRV), sleep hours/stages, daily steps, and stress scores
Biomarker cards displaying current metrics with risk indicators
HRV analysis (most critical stress biomarker): <10ms = critical, 10-15ms = high stress, 15-20ms = moderate
"Use Mock Data" fallback for demo purposes when real wearable data unavailable

📅 Smart Calendar Management

Week view with color-coded event types (caregiving, work, self-care, medical)
Add/edit/delete appointments with conflict detection
Overload detection: visual warnings for back-to-back appointments and 5+ daily events
AI-powered schedule optimization (Gemini-powered rescheduling suggestions)

🤖 AI Intervention Engine

Context-aware intervention generation using Google Gemini
Analyzes caregiver's actual calendar, tasks, biomarkers, and support network
Provides specific, actionable solutions with quantified time savings
Five intervention categories: delegation, automation, rescheduling, self-care blocking, respite care
Each intervention includes: problem statement, solution, time saved per week, difficulty level, and one-tap action

💬 Delegation Assistant

Contact management integrated with profile data
Task selection interface for delegation
AI-drafted messages that remove the emotional barrier of asking for help
Warm, specific, reciprocal message templates ready for SMS delivery
Planned: Direct SMS integration via Twilio for one-tap sending

📈 Provider Dashboard

Real-time monitoring dashboard at /dashboard endpoint
Displays all caregiver profiles with current burnout scores
Auto-refreshes every 60 seconds for continuous monitoring
HTML interface for healthcare providers, case managers, and family coordinators
Aggregate risk assessment across multiple caregivers

Demo Personas

We built three realistic caregiver personas with actual data:

Sarah (Critical - Score 84):

48-year-old marketing manager
2 kids + elderly mother with mobility issues
8+ weekly appointments, 14 pending tasks (8 overdue)
Sleep: 3.5-4 hours/night, HRV: 8-14ms (critical stress)
Demonstrates urgent intervention generation and delegation workflows

Maria (Moderate - Score 52):

45-year-old part-time nurse
1 teenage son + father with dementia
5-6 weekly appointments, 10 pending tasks
Sleep: 5-6 hours/night, HRV: 17-22ms (borderline)
Shows early detection and preventive interventions

Emma (Managing - Score 24):

42-year-old remote software developer
2 young kids, strong support network
4-5 weekly appointments, 8 current tasks (none overdue)
Sleep: 6.5-7.5 hours/night, HRV: 40-55ms (healthy)
Proves system doesn't over-alert for well-managed caregivers

How we built it

Architecture

Backend (FastAPI + Python)

RESTful API with FastAPI providing automatic OpenAPI documentation
Pure Python burnout calculation algorithm (no external scoring APIs for reliability)
Supabase PostgreSQL for data persistence with comprehensive schema
Google Gemini API integration for intervention generation and schedule optimization
Health data processing with JSON upload endpoints
Provider dashboard with HTML rendering and auto-refresh

Frontend (React Native + Expo)

Cross-platform mobile app built with Expo for rapid development
TypeScript for type-safe development
Zustand for lightweight state management
NativeWind for Tailwind-style native styling
react-native-chart-kit for burnout trend visualizations
Six main screens: Dashboard, Calendar, Wearable, Interventions, Delegation, History

Database (Supabase)

profiles: User demographics, care type, support contacts (JSONB)
calendar_events: Appointments with types and density flags
burnout_scores: Latest scores with 4-component breakdown
burnout_daily: Historical scores for trend visualization
wearable_data: HRV, sleep, steps, stress metrics by date
tasks: Pending and overdue tasks for load calculation
interventions: AI-generated action recommendations

Technical Implementation Highlights

4-Component Burnout Formula We implemented the official scoring algorithm from scratch in Python:

BurnoutScore = (WearableRisk × 0.30) +
               (CalendarTaskRisk × 0.30) +
               (SemanticRisk × 0.25) +
               (SelfReportRisk × 0.15)

Each component has detailed sub-calculations:

Wearable Risk: HRV (45%), Sleep (35%), Activity (10%), Stress (10%)
Calendar/Task Risk: Appointment density, back-to-back conflicts, overdue tasks
Semantic Risk: Burnout language detection from voice transcripts
Self-Report Risk: Mood trends and sleep quality

Real-Time Health Data Processing

JSON upload endpoint accepts multi-day health data batches
Upserts to wearable_data table with conflict handling
Automatically triggers burnout recalculation
Returns detailed processing statistics (inserted, updated, failed records)

AI Intervention Generation

Gemini API analyzes caregiver context (calendar, tasks, biomarkers, contacts, location)
Structured JSON output with specific solutions, not generic wellness tips
Quantified impact: "Saves 2.5 hours/week" for every intervention
Difficulty assessment (Easy/Moderate/Hard) for prioritization

Schedule Optimization

/optimize endpoint fetches current events and burnout score
Gemini proposes specific event time changes with reasoning
Returns structured proposed changes: event_id, new_start, new_end, reason
Frontend integration ready for one-tap acceptance

Challenges we ran into

1. Wearable Data Integration Complexity

Initial plan: Parse Garmin .fit binary files using fitparse library
Challenge: .fit file format has device-specific variations and complex field mappings
Solution: Implemented flexible JSON upload endpoint first for rapid prototyping, with .fit parsing as optional enhancement
Result: More flexible system that accepts data from any wearable brand

2. Burnout Scoring Without External APIs

Original design relied on WolframAlpha API for calculation validation
Challenge: External API calls added latency (2-5 seconds) and created demo failure risk
Solution: Implemented pure Python calculation with comprehensive unit tests
Result: Sub-500ms scoring, zero network dependencies, deterministic results

3. Gemini API Structured Output

Challenge: Getting consistent JSON structure from LLM for interventions
Early attempts returned inconsistent formats and generic advice
Solution: Refined prompt engineering with explicit JSON schema, caregiver-specific context injection, and few-shot examples
Result: Reliable structured outputs with specific, actionable interventions

4. Mobile State Management Across Screens

Challenge: Keeping burnout score, health metrics, and calendar data synchronized across 6 screens
React Native's navigation model made prop drilling unwieldy
Solution: Implemented Zustand for lightweight global state with persistence
Result: Instant data updates across all screens when any component changes

Accomplishments that we're proud of

🏆 First Platform to Combine Wearables + AI for Caregiver Burnout

No existing solution synthesizes objective physiological stress markers (HRV, sleep) with AI-driven intervention generation specifically for caregivers
We created a novel application of health technology that addresses an $873 billion invisible labor crisis

⚡ Production-Quality Burnout Algorithm

Implemented peer-reviewed burnout research (HRV as primary stress indicator) in a working system
4-component weighted formula with scientifically-grounded thresholds
Validated across 3 realistic personas with expected score differentiation (24, 52, 84)

🎨 Complete Full-Stack Mobile Experience

Built functional iOS/Android app in <24 hours with professional UI/UX
Six fully-wired screens: Dashboard, Calendar, Wearable, Interventions, Delegation, History
Smooth animations, chart visualizations, and intuitive navigation

🤖 Context-Aware AI That Actually Helps

Interventions reference actual calendar conflicts: "Your dentist appointment Wed 3pm conflicts with school pickup"
Quantified impact on every suggestion: "Saves 2.5 hours/week"
Delegation messages use caregiver's real contact names and tone preferences

📊 Provider Dashboard for B2B Scalability

Real-time monitoring view for healthcare systems and employers
Auto-refreshing aggregate dashboard shows all caregivers' status
Demonstrates business model viability beyond consumer app

🔧 Clean Architecture Under Time Pressure

Comprehensive API documentation with OpenAPI/Swagger
Modular backend with clear separation: routes, modules, database layer
TypeScript interfaces throughout frontend for type safety
Detailed inline documentation in PROJECT_GOALS.md, API.md, DATABASE.md

What we learned

Technical Lessons

Multimodal Data Fusion is Hard

Combining calendar events, wearable metrics, voice data, and self-reports requires careful normalization and weighting
Different data sources have different update frequencies and reliability levels
Learned to build robust systems with graceful degradation when data sources are incomplete

LLM Prompt Engineering for Structured Outputs

Generic prompts produce generic advice ("practice self-care," "get more sleep")
Specificity comes from injecting actual user context: calendar conflicts, contact names, biomarker values
JSON schema enforcement in prompts dramatically improves output consistency

Hackathon vs. Production Tradeoffs

Authentication, OAuth, and HIPAA compliance add weeks of work with zero demo impact
Pre-loaded personas and mock data enable smooth demos without network dependencies
Knowing when to cut scope is the most important technical decision

Domain Research

Caregiver Burnout Science

HRV (heart rate variability) is the most research-validated objective stress biomarker
Burnout is gradual (2-4 week decline), not sudden—early detection is key
Caregivers know they need help but face emotional barriers to asking—delegation assistance removes this friction

Healthcare Market Dynamics

Employers lose $17-33 billion annually to caregiver burnout (absenteeism, turnover)
B2B model ($50/employee/year) more scalable than direct-to-consumer
Provider dashboards align with existing case management workflows

Impact Measurement

Quantified time savings ("2.5 hours/week") resonates more than qualitative benefits
Caregivers need to see ROI to justify behavior change
Tracking intervention acceptance and outcome is crucial for product-market fit