Skip to content
Mai Trinh
AIReusable Product

Offline AI Chat App

A native Android application running on-device large language models completely offline. Delivers private, zero-latency generative AI interactions without internet connectivity or cloud API costs.

This is an independent software product developed with full technical ownership. Source code, customization, or private deployment are available.

Overview

Halo is a fully offline conversational AI mobile application built with Flutter. By running Qwen3.5-2B (quantized Q4_K_M) through llama.cpp (llamadart) and integrating sherpa-onnx with a Zipformer model for offline Vietnamese speech recognition, it generates answers completely on-device without servers, user accounts, or internet permissions.

Product Demo & Walkthrough

Demo videos coming soon — please check back or contact me for a live demo.

Key Features

  • 100% on-device local LLM execution using llama.cpp and Qwen3.5-2B (no network permissions requested)
  • Offline Vietnamese voice-to-text input powered by sherpa-onnx and Zipformer speech model
  • Optimized inference speed: cached system prompt state and 4 prefill threads reducing time-to-first-token to 0.93s
  • In-app local model file import with streaming SHA-256 verification and atomic installation
  • Deterministic pre-processing in Dart for calendar and arithmetic accuracy
  • Local chat history persistence with SQLite and state management with Riverpod

Best For

Privacy-Conscious Users

100% on-device AI inference with zero network traffic, ensuring complete conversation privacy.

Offline Learning & Productivity

Knowledge Q&A, writing assistance, and concept explanations functioning in airplane mode.

Internal Assistants Without Cloud Inference

Zero recurring third-party API subscription costs while keeping internal notes confidential.

Organizations With Limited Connectivity

Field operations in remote regions, ships, construction sites, and network-isolated zones.

Technology Stack

Technology Stack

FlutterDartRiverpodllama.cppQwen3.5-2Bsherpa-onnxSQLite

Supported Platforms

Android 10+ (Tested on Galaxy S10 & Tab S9 FE)

Customization & Delivery Possibilities

  • Integration of domain-specific GGUF quantized models (Qwen, Gemma, Phi)
  • Custom system prompts and specialized field-inspection assistant behaviors
  • Offline voice model customization for specific vocabularies and acoustic conditions
  • White-label branding and adaptation for mission-critical offline environments

Interested in Offline AI Chat App?

Reach out to request customization, ask questions about integration, or receive a live walkthrough.

Contact About This Product