← Back to Projects
CLI Tool

AI Survival

A fully offline, portable AI assistant that runs a local LLM from an external HDD/USB — no internet required after setup.

DockerShellllama.cppPythonLLM

Overview

AI Survival packages a complete local LLM stack onto an external HDD or USB drive. Plug it into any machine, run a single setup script, and chat with a quantized language model through a browser-based UI — fully air-gapped. It supports two deployment modes: Docker Compose (cross-platform, recommended) and a Python-native local mode for machines without Docker.

Key Features

  • One-command startup (setup.sh then start.sh) across Linux, macOS, and Windows
  • Two deployment modes: Docker Compose and local Python — no internet required after initial prep
  • Interactive model downloader with resume support (TinyLlama 1.1B → Llama 3.1 8B)
  • NVIDIA CUDA GPU acceleration via docker-compose.gpu.yml override
  • validate.sh pre-deployment checklist covering 30 checks across 6 categories
  • Offline Docker installer pre-caching and save-images.sh for fully air-gapped targets

Technical Decisions

Chose llama.cpp over Ollama for direct inference control and a built-in OpenAI-compatible HTTP API, with Open WebUI as the frontend since it natively speaks that API. Docker Compose handles service orchestration and volume-mounted GGUF models, while docker save/load enables true air-gapped deployment by exporting image tars to the HDD. Q4_K_M quantization was selected as the default — best quality-to-size ratio for resource-constrained portable hardware.