Taking a machine from a bare OS to a private inference API, with readable scripts and nothing stored.
Standing up local AI on hardware you own involves drivers, runtimes, storage, inference engines, and a series of machine-specific decisions. It is rarely one command.
ai-on-a-box probes the actual machine and determines what can run on it. It then produces readable scripts, plus three files the owner keeps: a profile, a plan, and a report. Every recipe carries an assurance tier that says how much evidence stands behind it, and there are no accounts and nothing stored.
One repeatable path from a bare OS to a private inference API on Linux, macOS, or Windows, on a GPU or the CPU. The catalogue holds seven recipes and an engine matrix covering Ollama, vLLM, Open WebUI, Whisper, Kokoro, and Stable Diffusion WebUI, and the same plan always produces byte-identical scripts.
The first hardware recipes target the machines we benchmark ourselves. On an RTX 3090, 33B and 34B models run entirely on the GPU at about 41 tokens per second; an ASUS GX10 runs a 120B model at about 13 tokens per second.
Book a free 30-minute call with the architect who would do the work, and we'll talk through what it would take.