July Hope

In disaster-prone regions and remote wilderness areas, cellular connectivity is often nonexistent, leaving people without access to essential survival information or navigation tools. Modern smartphones become useless bricks when there's no internet, yet they contain powerful hardware capable of running AI models and rendering maps offline. Current solutions either require internet connectivity, are too complex for non-technical users, or fail to provide integrated assistance. There is no single, easy-to-use application that combines offline AI-powered survival guidance with offline navigation. This gap puts lives at risk when every second counts. People need a tool that can provide step-by-step survival advice and navigation assistance without relying on fragile network infrastructure.

JulyHope addresses this critical gap by creating a completely offline, self-contained application that operates independently of internet connectivity. The solution comprises two core components working in harmony: an offline AI survival assistant and an offline navigation system. The AI component uses the Gemma-4-E2B-4.6B model, quantized and optimized for mobile devices, running entirely on-device using llama.cpp. This eliminates the need for cloud APIs, ensuring privacy and availability in any situation. The model provides clear, step-by-step survival guidance for scenarios like first aid, water purification, fire starting, and shelter building. GPU acceleration via Vulkan ensures reasonable response times even on mid-range devices. The navigation component pre-downloads OpenStreetMap tiles for selected regions (zoom levels 5-12), enabling complete offline map functionality with GPS tracking, place search, and coordinate-based navigation. The tile downloader uses policy-compliant sources (Thunderforest) with appropriate caching and rate limiting. The app uses Hive for local storage of conversations and map tiles, Riverpod for state management, and implements proper Android permissions. A comprehensive download manager handles the 2.29 GB model and map region downloads with progress tracking and screen wake lock to prevent interruptions. This integrated approach provides a lifeline in emergency situations where every second matters, democratizing access to survival knowledge and navigation tools regardless of connectivity. By combining AI and mapping in one intuitive interface, JulyHope empowers users to make informed decisions and navigate safely in any environment.

JulyHope is built on Flutter 3.24+ with Dart 3.3+, targeting Android API 26+ devices. The app uses Riverpod 2.5.1 for robust state management, implementing StateNotifier providers for chat, map, and download states. Hive 2.2.3 serves as the local NoSQL database, efficiently storing conversations with chat history and offline map tiles. For offline mapping, flutter_map 8.2.2 renders OpenStreetMap tiles with latlong2 for coordinate handling, while geolocator 14.0.2 provides GPS tracking and location services. Permission handling uses permission_handler 12.0.1, connectivity_plus monitors real-time network changes, and wakelock_plus ensures screen remains on during large downloads. Device detection uses device_info_plus for GPU capability identification. The app employs path_provider 2.1.5 for file system access and http 1.6.0 for secure model and tile downloads. Android permissions include internet, fine/coarse location, storage (Android 12 and below), and notifications (Android 13+). The UI is fully responsive across all mobile screen sizes.

The AI inference pipeline uses llama_flutter_android 0.2.6, wrapping llama.cpp for native on-device model execution, ensuring complete privacy and offline functionality. The model is Google's Gemma-4-E2B-It (4.6B parameters), quantized to IQ2_M format (2.29 GB) and sourced from HuggingFace. Architecture includes 35 layers, 1536 embedding size, 8 attention heads, 20 shared KV layers, and 131K context length. Vulkan GPU acceleration delivers 25-40 tokens/second. The model uses auto-detected Gemma-4 chat templates with tool-calling and reasoning channels. Download validation checks file integrity (minimum 2.29 GB, SHA256: 3d95ada2a122c9c0b42803317239b64b262ac9226a307ff895b3d87eec0c2acd), and response cleaning removes special tokens for clean output. Inference parameters include temperature 0.7, top-p 0.95, and mirostat sampling for quality.