🛡️ AEGIS: Autonomous Edge-Guided Interface Sentinel

A 100% Offline, Privacy-First Standalone Android OS Agent Powered by On-Device Gemma 4 & LiteRT-LM

---

Overview & Problem Statement

Modern personal AI assistants suffer from two critical flaws: cloud dependency (exposing sensitive personal conversations, screen context, and telemetry to remote servers) and lack of physical device autonomy (inability to navigate apps or execute multi-step OS workflows locally).

AEGIS transforms an Android smartphone into an autonomous, voice-driven personal sentinel operating entirely on-device without internet access, third-party cloud APIs, or external telemetry.

Key Architectural Innovations & Features

1. 100% On-Device LiteRT-LM Engine: Employs Google AI Edge LiteRT-LM with gemma-4-E2B-it.litertlm accelerating native LLM tool calling directly on the mobile GPU with zero network calls.
2. Offline Dual-Engine Voice Pipeline: Continuous 16kHz background wake-word listener (“Gemma”) powered by Vosk, featuring an atomic hardware-level microphone handoff to Android STT and floating Arc-Reactor HUD.
3. Autonomous OS Navigation & Accessibility: Real-time view hierarchy traversal (AccessibilityNodeInfo) and programmatic touch gesture dispatching (GestureDescription) across target applications (WhatsApp, Samsung Notes, Calendar, Settings).
4. Active Communication & Call Engine: Native tools to autonomously send WhatsApp/SMS messages (send_message), place direct calls or launch the dialer (call_handling), and inspect recent call logs.
5. Local Context & S.T.A.R.K. Second-Brain Sync:
   • Passive notification listener buffering WhatsApp and Gmail messages into local encrypted Room SQLite.
   • Built-in Notes memory sync (Samsung Notes / Keep) and calendar extraction via local CalendarContract.Instances.
   • Native scanner for voice recordings and local file searches.
   • Instant export of facts, notifications
6. Zero-Trust Biometric HITL Gate: Zero-trust risk classification (RiskAssessor) that intercepts destructive or sensitive operations (payments, deletions, credential fields) behind an Android Biometric Prompt.
7. Self-Improving Episodic Memory: ReAct tool chains are indexed into SQLite (SkillDao) enabling sub-second cached recall for recurring intent patterns.