For a while now I’ve been building my local AI stack on this network — a 64 GB DDR4 box with a 24 GB GPU, a local model server, a browser chat interface, and an agent layer that reads my own documents. None of it touches a cloud.
The hard part was never the hardware. It was the decision — what fits in your memory, which install path to take, which model to start with, and when a piece of gear is actually worth buying. I kept seeing people burn money on the wrong machine because nobody showed them the math first.
So I wrote it down. offlinestack.com is live today.
What it is
A practical field guide to building a private AI stack — from the machine you already own to a local agent that’s genuinely useful. Two editions: a $29 guide (read it) and a $49 Builder Bundle (build it tonight, with an editable starter pack and free updates while it’s a v1.x release). Everything is version-aware on purpose: it teaches the decisions and points you at current sources instead of freezing model tags that will be stale by next quarter.
Two things to push back on, if you’ve been hearing the other story
You don’t need new hardware. DDR4 remains a practical baseline. A 32 GB CPU-only machine is the floor; a last-generation box with 12–16 GB of VRAM is a real path. The reference build in the guide is a one-generation-behind machine — the same class of box I run mine on.
The setup is easier than the lore suggests. If you have a compatible NVIDIA card, TrueNAS Community Edition (free) ships an official app catalog where Ollama and Open WebUI are a deploy-and-go install. An afternoon, minimal terminal.
The scorecard is the part to try first
Seven questions, two minutes, and a straight answer about your hardware — including the honest one: not yet, and here’s the cheapest fix. It’s free, and it’s the front door to the site:
Take the Hardware Fit Scorecard
If it says you’re ready, the guide will show you the way. If it says you’re not, it’ll save you a wrong purchase — which, for a $29 product, is most of the value.
Private AI should feel useful, not fragile. That’s the whole pitch.
