Skip to content

Through the flames

I built it for my grandmother because she is bad with smartphones Elderly people lose billions to scams yearly, panic during emergencies, and struggle with basic apps.

Code Video Cerebral Valley

Core trick

Wake-word voice input routes to an LLM that drives the Android UI by tapping, scrolling, and typing on screen for the user.

Tags

accessibilityvoiceagents

Description

I built it for my grandmother because she is bad with smartphones Elderly people lose billions to scams yearly, panic during emergencies, and struggle with basic apps. Bigger fonts aren't the answer. HelpGranny doesn't teach grandma to use the phone, it uses the phone for her. She just talks. The AI taps, scrolls, types, and navigates like a grandkid sitting next to her. -Scam Shield — spots sketchy calls/texts and warns her. Out loud. -Emergency — one sentence, call placed to emergency contact -WhatsApp — reads messages, sends messages, all by voice. No UI to learn. Phone just listens and acts.
rafaym1/HelpGranny★ 20 forksPythonCreated Apr 25, 2026 · last push Apr 25, 2026

README excerpt

# HelpGranny

A mobile-use agent for voice-controlled assistant for elderly users that controls an Android phone on their behalf. The user speaks, and the phone acts.

<p align="center">
  <img src="docs/screenshot.jpg" width="300" alt="HelpGranny app screenshot"/>
</p>


## The problem

Elderly people lose billions to scams yearly, panic during emergencies, and struggle with basic apps. Bigger fonts aren't the answer. HelpGranny doesn't teach grandma to use the phone; it uses the phone for her. She just talks. The AI taps, scrolls, types, and navigates like a grandkid sitting next to her.

## What it does

- Listens for the wake word "Helper" continuously, or responds to a tap
- Routes the request to the right specialist agent (WhatsApp, scam detection, emergency, or a direct answer)
- Sees the screen using Android's Accessibility Service, no ADB or laptop needed
- Taps, scrolls, and types on the phone using real gestures
- Speaks the result back using on-device TTS

Example commands:
- "Send a WhatsApp to my son saying I had lunch"
- "Someone is calling saying they're from the bank and need my PIN"
- "I fell, get help"
- "What time is it?" (answered directly, no screen interaction needed)

## How it works

```
User speaks
    |
    v
Wake word / TAP TO TALK
    |
    v
Orchestrator (Claude Haiku) -- routes intent
    |
    |-- DIRECT       --> speak answer
    |-- WHATSAPP     --> agent loop
    |-- SCAM_SHIELD  --> agent loop
    |-- EMERGENCY    --> agent loop

Team