Caregiver-facing home safety, voice assistance, and device coordination for older adults.
Eldora connects a caregiver mobile app, a public web presence, a central backend, a voice-processing service, DoraBot, and the DoraShield wearable. The system collects device status and safety events, processes DoraBot voice requests, returns spoken responses, and gives caregivers access to alerts, reminders, device controls, and activity summaries.
Eldora is a support and coordination system. It is not a medical device, does not diagnose conditions or prescribe care, and does not replace emergency services, clinical assessment, or caregiver judgment.
Home safety workflows cross several interfaces: a wearable detects an impact, a companion device captures a request, cloud services route the event, and a caregiver needs enough context to respond. Eldora keeps those responsibilities separated while maintaining one end-to-end path from the home devices to the caregiver application.
- A caregiver signs in to the mobile app, creates or joins a home, and configures its devices.
- DoraBot is provisioned on the local network and paired with the caregiver's home through the backend. DoraShield exposes local Wi-Fi setup and a pairing token, but its current firmware requires backend pre-registration or a compatible device-authentication policy before production requests can succeed.
- Registered devices send heartbeats and telemetry; DoraShield can submit fall events when backend authentication accepts its device key, while DoraBot records voice and forwards audio through the backend.
- The backend authenticates devices, stores system state, forwards voice audio to the AI pipeline, and applies notification and command rules.
- The AI pipeline transcribes audio locally with Faster-Whisper, classifies intent with local rules, calls an OpenAI-compatible Qwen endpoint for response and emotion tasks, plans supported reminders, and synthesizes cached speech with Edge TTS.
- The backend records relevant results, queues DoraBot playback or other device commands, and sends caregiver notifications when its routing rules require them.
- The mobile app presents homes, devices, alerts, scenes, reminders, and analytics; the public web app provides product information, legal pages, and the current mobile download route.
| Repository | Responsibility |
|---|---|
| Eldora-Frontend | Next.js public site, mobile download redirect, privacy policy, and account-deletion guidance |
| Eldora-Mobile | Expo and React Native caregiver app for authentication, homes, pairing, devices, alerts, scenes, reminders, voice settings, and analytics |
| Eldora-Backend | Express API for users, homes, devices, IoT events, commands, notifications, scenes, voice orchestration, reminders, and persistence |
| Eldora-AI-Pipeline | FastAPI voice service for local transcription, Qwen-backed response and emotion tasks, local intent and reminder logic, and speech synthesis |
| Dorabot | ESP32-S3 companion firmware for provisioning, pairing, heartbeats, voice capture, audio playback, display output, and backend command polling |
| DoraShield | ESP32 wearable firmware for MPU6050 motion sensing, local Wi-Fi setup, local buzzer alerts, telemetry, fall-event reporting, and command polling; production backend authentication still requires integration work |
| Layer | Technology |
|---|---|
| Public web | Next.js 16, React 19, TypeScript, Tailwind CSS |
| Caregiver mobile | Expo 55, React Native 0.83, TypeScript, Expo Router, TanStack Query, Zustand, Firebase |
| Backend | Node.js, Express 5, TypeScript, Prisma 7, PostgreSQL, Firebase Admin, Zod |
| Voice pipeline | Python 3.11, FastAPI, Faster-Whisper, httpx, OpenAI-compatible Qwen, Edge TTS |
| DoraBot | ESP32-S3, Arduino C++, I2S audio, LovyanGFX, HTTP, ArduinoJson, NVS |
| DoraShield | ESP32, Arduino C++, MPU6050, I2C, HTTP, ArduinoJson, NVS |
- Caregivers remain responsible for action — Eldora surfaces events and requests but does not make clinical decisions.
- The backend coordinates shared state — device identity, homes, alerts, commands, reminders, and histories are managed centrally.
- Devices keep focused responsibilities — DoraShield detects and reports motion events; DoraBot handles the in-home voice and display interface.
- Safety routing does not depend only on generated text — intent checks and caregiver-notification rules also exist outside the remote language-model call.
- Degradation is explicit — the voice service can return local fallback responses when Qwen is unavailable, while health reporting marks incomplete configuration as degraded.
- Configuration stays environment-specific — credentials and deployment settings are not embedded in organization documentation.
- DoraShield uses threshold-based motion detection. A reported event is a safety signal, not a confirmed medical fall assessment.
- The AI pipeline transcribes speech, generates short responses, estimates an emotion label, classifies supported intents, and plans supported reminders. It does not diagnose, prescribe, contact emergency services, or directly own caregiver notifications.
- The backend decides how device events and voice requests become persisted records, commands, or notifications.
- The mobile app is the caregiver interface; it is not an autonomous monitoring or emergency-response service.
- The public frontend does not contain the authenticated caregiver workflow.
The active repositories currently cover the public website and APK download route, caregiver authentication and home management, DoraBot provisioning and pairing, DoraBot voice capture and playback, DoraShield local setup and motion detection, backend device and notification orchestration, scene automation, voice and emotion records, caregiver analytics, and elder voice reminders. DoraShield production authentication remains an integration boundary rather than a completed end-to-end flow. Repository code and deployment configuration define the implemented scope; roadmap items are not treated as available features.