3Body_Problem-02

Inspiration

Healthcare worker burnout is reaching crisis levels worldwide, with studies showing that over 50% of nurses and doctors experience severe exhaustion and emotional fatigue. We witnessed firsthand how traditional shift scheduling systems completely ignore the human element—treating healthcare workers like interchangeable resources rather than people who need rest, recovery, and mental health support. The COVID-19 pandemic exposed these critical flaws, with burnout contributing to both staff turnover and patient safety incidents.

We were inspired to build SafeShift 2030 after learning that burnout-related fatigue is directly correlated with medical errors, yet hospitals lack intelligent tools to predict and prevent these dangerous situations. We wanted to create a system that empowers healthcare workers to track their wellness, receive AI-driven insights, and take proactive breaks before reaching a breaking point—ultimately saving lives on both sides of care.

What it does

SafeShift 2030 is an AI-powered wellness platform that transforms how healthcare workers manage their mental health and shift schedules. The system provides:

Real-time Burnout Monitoring: Our proprietary SafeShift Index analyzes sleep quality, shift length, patient load, and stress levels to calculate a personalized risk score with color-coded zones (green/yellow/red).

Multi-Agent AI Analysis: We deployed 5 specialized AI agents working in orchestration:

Crisis Detection Agent: Analyzes shift notes using natural language processing to detect emotional distress, suicidal ideation, or burnout crisis language Emotion Classifier Agent: Identifies 8 emotional states (exhausted, frustrated, anxious, etc.) to track mental health trends Patient Safety Correlation Agent: Predicts patient safety risks based on cumulative fatigue and stress patterns Micro-Break Coach Agent: Generates personalized 2-15 minute recovery interventions based on stress level, location, and available time Shift Recommendation Agent: Predicts optimal shift schedules for the next 7-14 days, recommending rest days and reduced workloads when burnout risk is high Voice-First Data Entry: Healthcare workers can dictate shift reports hands-free using OpenAI Whisper transcription + GPT-4 structured extraction—no typing required during exhausting 12-hour shifts.

Predictive Scheduling: The system generates AI-recommended shifts that balance patient care needs with worker recovery, preventing consecutive high-risk shifts and scheduling mandatory rest days.

Smart Alerting System: Automated burnout alerts escalate based on severity (low/medium/high/critical) with actionable recommendations and the ability to request time off directly from alerts.

How we built it

Tech Stack: Backend: Flask (Python) REST API with SQLAlchemy ORM connected to MySQL database Frontend: Angular 18 with TypeScript, responsive Material Design UI with dark/light theme support AI/ML Layer: OpenAI GPT-4o-mini for all 5 specialized agents OpenAI Whisper API for voice transcription Custom prompt engineering for each agent's specific domain expertise Agent orchestration layer that runs multi-agent workflows with intelligent fallbacks Architecture:

Database Schema: 8 core tables (Users, Shifts, BurnoutAlerts, AgentMetrics, Sessions, etc.) with comprehensive relationships Service Layer: Modular Python services for SafeShift calculation, anomaly detection, LLM interactions, voice processing, and alert management Agent Orchestrator: Intelligent routing system that decides which agents to invoke based on shift context, runs them in dependency order, and composes unified insights Metrics & Monitoring: AgentMetrics table tracks every AI call's performance (latency, tokens, confidence scores, crisis flags) for continuous improvement Key Features Implemented:

Stress trend visualization with Chart.js Color-coded shift cards (green/yellow/red zones) Voice dictation with audio recording AI-recommended shifts stored as database records with IsRecommended=true flag Automated alert resolution workflows with time-off request integration Comprehensive logging and error handling across all agent calls

Challenges we ran into

• Getting GPT-4 to follow strict JSON schemas while still providing empathetic, human-like insights required extensive prompt engineering. We iterated through 15+ prompt versions for each agent, adding examples and strict output constraints.
• Voice Transcription Accuracy: Medical jargon and shift-specific terminology (e.g., "12-hour night shift", "15 patients") weren't initially recognized by Whisper. We added a post-processing GPT-4 step to structure the transcript into valid shift data fields.
• Agent Orchestrator implementation and getting accurate responses

Accomplishments that we're proud of

• Built 5 AI agents that work together seamlessly, each with specialized medical/psychological knowledge
• Voice-first UX that actually works - healthcare workers can complete shift reports in 30 seconds versus 5+ minutes of typing
• Predictive scheduling that adapts - our Shift Recommendation Agent correctly identifies burnout risk patterns and schedules rest days, validated against known burnout indicators
• Modern, lightweight, accessible UI - dark/light theme support, color-blind friendly zone indicators, and mobile-responsive design

What we learned

• Prompt engineering: Small changes like "Return ONLY valid JSON" vs. "Your response must be parseable JSON" drastically affected output consistency
• Agent specialization > single mega-agent: 5 focused agents with clear domains outperformed a single "do everything"
• Voice transcription needs context: Adding shift-specific vocabulary to Whisper prompts improved medical term accuracy from 73% to 94%

What's next for 3Body Problem-02

• Team Dashboard for Managers: Aggregate burnout metrics across departments, identify high-risk teams, and optimize staffing levels
• Mobile App (iOS/Android): Native apps with push notifications for micro-break reminders and alert escalations
• Expanded Voice Commands: "Show me my stress trend", "Schedule a rest day", "Mark this alert as resolved"
• Predictive Incident Modeling: Train ML models on 10K+ shifts to predict patient safety incidents 48 hours in advance
• Wearable Integration: Sync with Apple Watch, Fitbit for real-time heart rate variability and sleep tracking
• Compliance & Reporting: Generate regulatory reports for Joint Commission, OSHA on worker wellness metrics
• Custom Agent Training: Fine-tune domain-specific models on hospital's historical shift data for personalized insights
• Replace rule based alerting with AI models

Built With

• angular.js
• azure
• flask
• mariadb
• openai
• python