CPR practice anywhere. A pressure-sensitive pad, a real-time AI patient simulator, and a voice that corrects you within 50 milliseconds.

Built for HackaBull VII (Apr 26, 2026).

Out-of-hospital cardiac arrest survival without bystander CPR is around 10%. With proper compressions, it doubles. The bottleneck is training — not motivation. Revive trains anyone, anywhere, with $50 of hardware and a browser.

You compress a force-sensitive pad. The pad streams pressure peaks over Web Serial. The web app:

1. Generates a fresh emergency scenario via Gemini 2.5 Flash (patient demographics, location, decision tree).
2. Simulates patient physiology in real time — heart rate, BP, rhythm, oxygenation — based on your compression depth and rate.
3. Coaches you out loud through three ElevenLabs voices — a calm instructor for the welcome, a 911 dispatcher reading scenario decisions, a panicked bystander with scenario-customized text.
4. Scores you against AHA protocol (depth 2.0–2.4", rate 100–120 BPM, recoil quality, decision timing).

Voice replaces the screen, so your eyes stay on the patient.

1. Cinematic intro → click anywhere → calm instructor voice welcomes
2. Click BEGIN CPR PROTOCOL → Gemini generates scenario → bystander panics with scenario-specific lines
3. Decision screen: 20-second window, pick first action (CPR / pulse check / call 911 / AED)
4. Compression screen: depth bar, BPM counter, rhythm wave, anatomical silhouette pulsing with rate, dispatcher voice barking corrections per batch ("Push harder", "Slow down", "Allow recoil")
5. Patient reaches ROSC or flatlines → debrief with performance index, breakdown, retry option

git clone https://github.com/VickyAnaJ/Revive.git
cd Revive/web
npm install
cp .env.local.example .env.local   # add ElevenLabs + Gemini keys
npm run dev
# open http://localhost:3000 in Chrome

For the hardware: open firmware/revive_firmware.ino in Arduino IDE, flash to UNO R4 Minima, plug into USB. The web app auto-detects via Web Serial. No Arduino? Press Option+Shift+Space for keyboard fallback.

Required env vars:

NEXT_PUBLIC_GEMINI_API_KEY=...
NEXT_PUBLIC_ELEVENLABS_API_KEY=...
NEXT_PUBLIC_ELEVENLABS_VOICE_ID_INSTRUCTOR=...
NEXT_PUBLIC_ELEVENLABS_VOICE_ID_DISPATCHER=...
NEXT_PUBLIC_ELEVENLABS_VOICE_ID_BYSTANDER=...

1. Plug Arduino UNO R4 Minima into USB.
2. cd revive/web && npm run dev.
3. Open http://localhost:3000 in Chrome (Web Serial requires Chromium).
4. Click anywhere to unlock audio. Click Connect Pad (or press Option+Shift+Space to fall back to keyboard mode).
5. Click Begin CPR Protocol to start a session.

If the pad is unplugged or the user is on Safari/Firefox, keyboard mode lets every spacebar press simulate a compression peak. Same scoring, same coach voice.

1. Ctrl+C in the terminal running npm run dev.
2. Unplug the Arduino.
3. Browser tab can be closed without cleanup — no persistent state outside .env.local.

If a dev server orphans (port 3000 stuck): pkill -f "next dev" clears it.

• Hardware: Arduino UNO R4 Minima + FSR pressure pad → JSON-over-Web-Serial
• Web: Next.js 16 + React 19 + TypeScript + Tailwind 4
• AI: Gemini 2.5 Flash (3 agents — patient simulation, coach phrases, scenario generation)
• Voice: ElevenLabs eleven_flash_v2 streaming TTS + pre-rendered MP3s + browser SpeechSynthesis fallback
• Tests: Vitest, 222 tests, full TypeScript

revive/
├── firmware/    Arduino sketch (force sensor → JSON peaks at 100Hz)
├── web/         Next.js app — UI, state machine, AI agents, voice pipeline
├── api/         FastAPI scaffolding (currently unused; agents run client-side)
├── contracts/   Shared JSON schemas
└── docs/        Internal design + audit docs (gitignored, kept private)

• Gemini 2.5 Flash generates scenarios, simulates patient physiology, and produces coach phrases. Three independent agents with structured JSON output, schema validation, and a 3-tier fallback (retry → cached scenarios → rule-based vitals).
• ElevenLabs flash_v2 drives all spoken voice. Three distinct character voices for instructor / dispatcher / bystander. Pre-rendered MP3s for instant coach corrections; streaming for scenario-specific lines.

When something fails on stage, the demo keeps working:

• Pad disconnected → keyboard fallback (Option+Shift+Space)
• Gemini times out → cached scenarios + rule-based vitals
• ElevenLabs 429 / network down → browser SpeechSynthesis
• Audio blocked → text overlays preserve every coach correction

Built at HackaBull VII at the University of South Florida.