OmniAgent: An Offline LLM-Based Intelligent Security Monitoring Platform for Android

OmniAgent is an offline AI security monitoring application for Android devices, with core features as follows:

1. Fully Offline AI Analysis: Implements local LLM inference via NDK/C++ runtime and Llama.cpp, no network dependency, protecting user privacy;
2. Real-Time Security Monitoring: Integrates multi-layer monitoring modules covering UI elements, system notifications, background processes, etc;
3. Clean Architecture Design: Layered architecture ensures maintainability and supports function expansion;
4. Privacy-First: All data processing is done locally on the device, eliminating cloud leakage risks. This project pioneers the offline intelligent paradigm in the mobile security field, suitable for privacy-sensitive users, network-restricted environments, and enterprise deployment scenarios.

With the popularity of mobile devices, security and privacy protection have become core user needs. Traditional security applications have two major pain points:

• Cloud Dependency: Data upload to the cloud brings privacy risks and is limited by network stability;
• Difficulty Deploying LLM on Mobile: The powerful reasoning capabilities of large language models are hard to run efficiently on resource-constrained mobile devices. OmniAgent emerged to introduce offline AI capabilities into the Android security field, enabling local intelligent monitoring to solve the above problems.

Local Neural Engine

Optimized via NDK/C++ to support local inference of GGUF format models, using the Llama.cpp framework to adapt to mobile resources, and can work normally in flight mode.

Clean Architecture Design

• UI Layer: Modern interface built with Jetpack Compose;
• Business Logic Layer: Use cases coordinate data layer operations;
• Data Layer: Room database persistence + neural engine communication.

Multi-Layer Monitoring System

1. Neural Shield: Scans UI elements based on Accessibility Service to alert phishing/malicious patterns;
2. Signal Watch: Listens to system notifications and blocks sensitive leaks or threat messages;
3. Omni Guardian: Foreground service monitors system health and background processes;
4. AI Inference Visualization: Dynamically displays the inference process and threat levels.

Development Language and Framework

• 100% Kotlin development, with coroutines supporting asynchronous concurrency;
• Jetpack Compose builds responsive UI to reflect system status in real time.

AI Inference Engine

Hybrid architecture: C++ layer handles high-performance inference, Chaquopy (Python) processes model loading/preprocessing/postprocessing.

Data Persistence

Room encrypts and stores scan records and security logs, providing type-safe SQL operations.

Background Scheduling

WorkManager implements periodic system audits to ensure security checks can be performed even when the app is not running.

Core Features

• Offline AI Capability: Download quantized models locally, all analysis is completed on the device;
• Real-Time Threat Detection: Accessibility Service monitors screen content to identify new threats;
• Intelligent Notification Analysis: Semantic understanding blocks fraud/leakage risks;
• Material3 Design: Dynamic dashboard displays security status and visualizes the inference process.

Application Scenarios

1. Privacy-Sensitive Users: Data remains local, eliminating cloud leakage;
2. Network-Restricted Environments: Security protection is still available without a network;
3. Enterprise Deployment: Custom models and policies to control sensitive data.

Mobile Resource Constraints

Reduce resource usage through model quantization, memory mapping, and chunked inference to balance performance and efficiency.

Battery Life Optimization

Intelligent scheduling + event-driven + WorkManager battery awareness to minimize power consumption.

Permission and Privacy Balance

Transparent data processing policy + local execution guarantee + open-source code to build user trust.

OmniAgent proves the feasibility of offline LLM in the mobile security field, with core values in privacy protection and no network dependency. Future plans include:

• Expanding to smart assistants, content moderation, and other scenarios;
• Continuously optimizing model compression and mobile adaptation;
• Relying on the MIT open-source license, welcoming community contributions of models and functions. This project provides a local-first intelligent solution for the mobile security ecosystem, aligning with the trend of increasing user privacy awareness.