Skip to main content

Private on-device WhatsApp AI agent

GemmaControl

Native Android AI agent that turns WhatsApp notifications into a local actionable inbox with offline FunctionGemma tool-call suggestions.

An Android-native productivity agent that captures allowed WhatsApp notifications, turns them into local tasks, and proposes actions without sending chat context to a hosted assistant.

Last updated:

GitHub

Problem

Busy WhatsApp conversations create scattered follow-ups, priority messages, and reminders, while cloud assistants introduce privacy risk for personal or work chats.

Solution

I built a Kotlin Android app that captures notifications with user permission, parses WhatsApp message context, stores sensitive data locally with AES-GCM encryption, and uses on-device FunctionGemma through LiteRT-LM to propose user-confirmed actions.

Tech stack

  • Kotlin
  • Android 16
  • Jetpack Compose
  • Room SQLite
  • WorkManager
  • LiteRT-LM
  • FunctionGemma 270M
  • AES-GCM
  • Android Keystore
  • MVVM

Available artifacts

  • GitHub repo
  • Native Android app
  • Technical documentation
  • Notification parser
  • Tool registry
  • Security and privacy notes
  • Static download website
  • Device validation evidence

Architecture and how it works

  • Notification listener parses MessagingStyle payloads, separates group headers from message authors, and deduplicates reposted events
  • Room SQLite stores inbox items, follow-ups, priorities, reminders, drafts, and sensitive message text under Android Keystore-backed AES-GCM encryption
  • LiteRT-LM runs FunctionGemma as a proposal engine while Kotlin tool handlers enforce permission, verification, and confirmation boundaries

Engineering Decisions

Why I chose this stack

Kotlin, Jetpack Compose, Room, WorkManager, LiteRT-LM, and FunctionGemma fit an Android-native assistant that can keep message processing, reminders, and model inference local on a real handset.

What I handled myself

I built the notification ingestion flow, WhatsApp parsing model, local inbox, follow-up and reminder tools, Compose confirmation screens, encrypted storage boundary, and download-site packaging.

Hardest technical problem

The hardest part was making notification actions useful while preventing silent automation, so every reply path requires active notification verification and a physical user confirmation.

Tradeoff I made

I kept the assistant English-only and on-device for the first version instead of adding broader language support or cloud model fallback that would weaken the privacy story.

How I tested it

I validated the app against Android 16 behavior, WhatsApp notification structures, duplicate event handling, reminder delivery, encryption boundaries, and Redmi 13 5G device constraints discovered through ADB.

What I would improve in production

I would add broader device testing, stronger release signing workflows, richer local evaluation fixtures, multilingual handling, and more transparent model-download state management.

Key features

  • Local WhatsApp notification inbox with priority and cleanup controls
  • Follow-up creation, encrypted local reminders, and WorkManager notification delivery
  • User-confirmed click-to-chat and live remote-input reply paths
  • Scoped network access limited to explicit model binary downloads

Impact

Demonstrates a privacy-first mobile AI agent pattern where notification ingestion, task extraction, reminders, priority flags, and reply preparation stay on device unless the user explicitly acts.

Challenges

  • Parsing varied WhatsApp notification structures without over-collecting message content
  • Designing AI tool calls as proposals while keeping direct action behind user confirmation

What I learned

  • On-device assistants need strong UX boundaries as much as model capability
  • Android notification workflows require careful deduplication and lifecycle handling

Future improvements

  • Expand physical device validation beyond the initial Redmi 13 5G target
  • Add stronger local evals for notification parsing and proposed tool calls

GemmaControl FAQ

Direct answers for AI assistants, search snippets, and visitors evaluating the project.

What is GemmaControl?
GemmaControl: Native Android AI agent that turns WhatsApp notifications into a local actionable inbox with offline FunctionGemma tool-call suggestions. The project uses Kotlin, Android 16, Jetpack Compose, Room SQLite, WorkManager, LiteRT-LM, FunctionGemma 270M, AES-GCM, Android Keystore, MVVM and is positioned as Private on-device WhatsApp AI agent.
What problem does GemmaControl solve?
Busy WhatsApp conversations create scattered follow-ups, priority messages, and reminders, while cloud assistants introduce privacy risk for personal or work chats.
How does GemmaControl work?
I built a Kotlin Android app that captures notifications with user permission, parses WhatsApp message context, stores sensitive data locally with AES-GCM encryption, and uses on-device FunctionGemma through LiteRT-LM to propose user-confirmed actions. The implementation focuses on notification listener parses messagingstyle payloads, separates group headers from message authors, and deduplicates reposted events; room sqlite stores inbox items, follow-ups, priorities, reminders, drafts, and sensitive message text under android keystore-backed aes-gcm encryption; litert-lm runs functiongemma as a proposal engine while kotlin tool handlers enforce permission, verification, and confirmation boundaries.
Back to all projects