A daily cognitive companion that disguises neuropsychological assessment as a five-minute morning conversation. Built for the H2AI 2026 Hackathon.
Earlybird tracks longitudinal speech biomarkers — recall, fluency, semantic coherence — and surfaces early cognitive decline signals to physicians before a crisis, not after.
Three portals, one patient (Eleanor M., 74F):
| URL | Who sees it | What it shows |
|---|---|---|
/ |
Everyone | Landing page |
/onboarding/eleanor_001 |
Caregiver | Patient profile setup |
/voice/eleanor_001 |
Patient | Morning voice check-in with Birdie |
/doctor/eleanor_001 |
Physician | 90-day cognitive trend chart + AI-generated SOAP note |
/family/sarah_001 |
Adult daughter | Plain-English weekly digest |
The demo moment: On the doctor dashboard, click the red dot on March 15 — Eleanor said "hospital" four times unprompted during a picture description task.
- Python 3.11+
- Node.js 18+
- A Gemini API key
- An ElevenLabs API key (optional — for natural voice; falls back to browser TTS if not set)
Create a .env file in the project root:
GEMINI_API_KEY=your-gemini-key-here
ELEVENLABS_API_KEY=your-elevenlabs-key-here # optional
cd backend
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt
python seed.py # Creates eleanor.db with 100 days of demo data
uvicorn main:app --reload # Runs on http://localhost:8000cd frontend
npm install
npm run dev # Runs on http://localhost:5173Open http://localhost:5173 to see the landing page.
| Layer | Technology |
|---|---|
| Frontend | Vite + React 18, React Router v6, Recharts |
| Backend | Python 3.11, FastAPI, Uvicorn |
| Database | SQLite (zero-config, pre-seeded, no network needed) |
| AI — Conversation | Gemini 2.0 Flash (fast, warm conversational turns) |
| AI — Assessment | Gemini 2.5 Pro (clinical biomarker scoring) |
| AI — SOAP Notes | Gemini 2.5 Pro (streaming clinical documentation) |
| Voice (TTS) | ElevenLabs custom voice / Web Speech API fallback |
| Voice (STT) | Web Speech API (Chrome/Edge) |
earlybird/
├── frontend/
│ └── src/
│ ├── pages/
│ │ ├── LandingPage.jsx # /
│ │ ├── Onboarding.jsx # /onboarding/:id
│ │ ├── VoiceGame.jsx # /patient/:id
│ │ ├── DoctorDash.jsx # /doctor/:id
│ │ └── FamilyView.jsx # /family/:id
│ └── App.jsx # React Router setup
├── backend/
│ ├── main.py # FastAPI routes
│ ├── database.py # SQLite schema + connection
│ ├── seed.py # Generates 100-day demo data
│ └── requirements.txt
├── .env.example
├── .gitignore
└── README.md
Built on the ADReSS Challenge 2020 (Luz et al., INTERSPEECH 2020) — automated speech analysis distinguishes Alzheimer's from healthy controls with 75.6% accuracy. The key biomarkers (type-token ratio, fluency count, repetition frequency, semantic coherence) manifest in daily conversation, not just clinical settings.
Earlybird's 4-turn structured conversation mirrors the Cookie Theft picture description protocol from the DementiaBank corpus, with an embedded delayed-recall task to measure short-term memory across a conversational distraction.
A caregiver email alert fires automatically when:
- The session is explicitly flagged
redby the AI assessment, or - The composite cognitive index drops ≥ 20% relative to the patient's 14-day rolling average
- Database: Migrate to Supabase Postgres with Row Level Security
- Transcription: Whisper API for higher-accuracy STT
- EHR integration: Epic FHIR API for SOAP note import
- Reimbursement: CPT 99483 (cognitive assessment) + RPM billing pathway