Echoes

Echoes is an adaptive care companion for people who need calm, familiar support through the day. A caregiver uploads a care plan PDF, the system turns it into a personalized experience, and the patient sees a simple, reassuring interface that changes with their needs.

Click to watch Echoes in action.

It combines:

• CopilotKit + AG-UI for live agent-to-UI streaming
• A2UI for declarative generative UI cards
• FastAPI for the Python agents
• Next.js for the app shell
• Linkup for research and music discovery
• ElevenLabs for calming voice output
• Google Gemini 3.5 Flash as the frozen model behind the agents

Tags

generative-ui a2ui copilotkit ag-ui nextjs fastapi langgraph gemini elevenlabs linkup accessibility healthtech hackathon

Description

This project is built for a London hackathon demo, but the idea is practical: one uploaded care plan becomes a tailored patient screen, a family monitoring view, and a caregiver research assistant.

The core flow is:

1. Caregiver uploads a PDF care plan on /setup
2. The setup agent extracts a structured patient profile
3. The patient page renders one A2UI card at a time on /patient
4. Panic mode switches the interface into a calming flow with voice and music
5. The family page shows activity history and live status on /family
6. The research page provides cited guidance on /research

Screenshots

Landing screen

Patient experience

Family monitoring

Features

• Care plan PDF upload and profile extraction
• Personalized patient onboarding and adaptive home screen
• One-card-at-a-time A2UI rendering for better cognitive accessibility
• Memory prompts, daily tasks, and medication reminders
• Panic mode with calming language and voice playback
• Music discovery for reassurance and grounding
• Research assistant with cited guidance
• Family dashboard with activity log and live status

Tech Stack

• Frontend: Next.js 15, React 19, TypeScript
• Backend: FastAPI, Python, LangGraph
• Agent transport: CopilotKit v2, AG-UI
• UI system: A2UI
• AI services: Google Gemini 3.5 Flash, Linkup, ElevenLabs
• Utilities: pdfjs-dist, Redis support, Zod

Project Structure

src/
├── app/
│   ├── setup/page.tsx
│   ├── patient/page.tsx
│   ├── family/page.tsx
│   └── research/page.tsx
├── a2ui/
│   ├── theme.css
│   └── catalog/
├── components/
└── lib/

agent/
├── main.py
└── src/
    ├── patient_agent.py
    ├── research_agent.py
    ├── setup_agent.py
    └── catalog.py

Getting Started

pnpm install
cp .env.example .env.local
pnpm doctor
pnpm dev

For local development without external model calls:

OFFLINE=1 pnpm dev

Environment Variables

Create a .env.local file with the keys you need:

GEMINI_API_KEY=your_key_here
LINKUP_API_KEY=your_key_here
ELEVENLABS_API_KEY=your_key_here

PATIENT_AGENT_URL=http://localhost:8123/patient
RESEARCH_AGENT_URL=http://localhost:8123/research
SETUP_AGENT_URL=http://localhost:8123/setup

Commands

Command	Purpose
pnpm dev	Start Next.js and the Python agent together
pnpm dev:web	Start only the Next.js app
pnpm agent	Start only the Python agent
pnpm doctor	Run the environment preflight check
pnpm smoke	Run the full smoke test gate
pnpm typecheck	TypeScript typecheck
pnpm test	Run runtime tests
pnpm validate-widget <path>	Validate an A2UI component

Demo Flow

1. Open /setup and upload the demo care plan PDF.
2. Review the extracted profile and approve it.
3. Open /patient to show the personalized briefing and memory cards.
4. Trigger panic mode to demonstrate calming voice and music.
5. Open /research and ask a clinical or family support question.
6. Open /family to show the live activity log and status panel.

Notes

• The patient experience is intentionally simple, warm, and low-friction.
• The A2UI catalog is the source of truth for all patient-facing cards.
• The screenshots in /img are meant to be committed and displayed in the README.
• If a layout issue appears in an A2UI component, run pnpm validate-widget <path> and then pnpm smoke.

Sponsor Footer

Built for the London A2A + Generative UI hackathon with:

• CopilotKit + AG-UI for the live agent stream
• A2UI for declarative generated UI
• Linkup for research and music discovery
• ElevenLabs for calming voice output
• Google Gemini 3.5 Flash as the model behind the agents