Sentri AI - AI Voice Companion for Senior Citizens
Link to open source: https://github.com/Abinayamohana/sentri-AI
SentriAI — Intelligent, Privacy-First Eldercare Companion & Safety Monitor
SentriAI bridges warm daily companionship and reliable, instant emergency protection for elderly users.
What it does
-
Companion: A warm, full-duplex conversational voice assistant powered by Agora Conversation AI (via the Agora Voice/RTC SDK) that keeps the user company and checks in on them naturally.
-
Reminder: On-device scheduling that guides users through daily medication, meals, and appointments.
-
Emergency: Offline, on-device AI detects falls and verbal distress and instantly dispatches the emergency pipeline.
-
Failsafe Mantra: If an elder in distress can’t explain their situation, repeating a simple predefined phrase (e.g., “Mimi Mimi Mimi” or “Mime Mime Mime”) 3+ times in a row bypasses the classifier and triggers the emergency pipeline.
Emergency pipeline
SentriAI lets the user configure an emergency helper contact (name and mobile number). When an emergency event is detected — a fall, verbal distress, or the failsafe mantra — the pipeline triggers automatically and sends a help message to that contact, including the user’s current location.
Tech stack
-
Frontend & Client: Native Android app in Kotlin + Jetpack Compose.
-
Real-Time Communications: Agora Voice/RTC SDK for low-latency, full-duplex voice streams to the Agora Conversational AI Cloud Engine.
-
Local Speech-to-Text: Whisper-tiny.en on PyTorch ExecuTorch for fast, local transcription.
-
Local LLM Classification: Gemma 3 1B running locally via the Google MediaPipe LLM Inference SDK.
-
Local Storage: Android Room Database for schedule caching and alert history.
-
Backend Orchestration: FastAPI (Python) for serverless WebRTC token minters and Agora Studio agent configuration.
Why it matters: Companion conversations run through Agora Conversation AI for rich, low-latency dialogue, while safety detection runs entirely on-device — so emergency protection is private, instant, and includes the user’s location the moment help is needed.
This build was uploaded as a hackathon project
