Off Grid: Utilizing Phone Hardware for Offline AI Applications
Off Grid is an open-source application designed to utilize the powerful GPU in your smartphone to perform AI tasks offline, without relying on cloud services. This includes text generation using llama.cpp at 15-30 tokens per second with any GGUF model, image generation with Stable Diffusion running NPU-accelerated on Snapdragon NPUs taking 5-10 seconds per image, voice transcription with Whisper, and vision AI with models like SmolVLM and Qwen3-VL.
The app operates entirely offline, ensuring no data leaves your device, which is particularly crucial in scenarios with no internet access, such as flights, in countries with censorship, or in compliance-heavy environments like hospitals. For Android, the app is available as an APK from GitHub, while iOS users need to build the app from source code.
To run Off Grid, Android users should use npm install followed by cd android && ./gradlew clean && cd .. and npm run android. iOS setup requires cd ios && pod install && cd .. and npm run ios. Ensure Node.js 20+ and Xcode 15+ for iOS, or JDK 17 and Android SDK 34 for Android are installed. This type of offline capability is particularly useful for developers concerned about data privacy and those working in internet-restricted or data-sensitive environments.
📖 Read the full source: HN AI Agents
👀 See Also
Exploring LiveDocs: An AI-native Data Analysis Notebook
LiveDocs offers a reactive notebook environment allowing data teams to perform multi-step analyses and maintain analysis end-to-end with the help of an AI agent.
Free pharmacovigilance signal detection tool built with Claude Code
A developer used Claude Code to build a free pharmacovigilance platform that analyzes 2.9 million FDA adverse event reports, running statistical signal detection that commercial platforms charge $50K-500K/year for. The tool is hosted for free on HuggingFace.
Open-source trust scoring hook for Claude Code monitors sessions, blocks protected paths
A developer built a Python hook that scores every Claude Code session on reliability, scope, and cost dimensions, blocks access to protected paths like .env files, and hash-chains events for tamper detection. The single-file tool is available on GitHub.
Prism MCP v5.1 adds 10x memory compression and agent learning from corrections
Prism MCP v5.1 introduces 10x memory compression via TurboQuant ported to TypeScript, enabling millions of memories on a laptop without vector databases. The update adds agent learning from user corrections and a visual knowledge graph interface.