THE PRIVATE AI STARTER STACK / VERSION 2.1
Choose a sensible starting model. Get local chat working. Then put it to work on your own notes—with concrete steps and checks along the way.
Ubuntu or existing TrueNAS · CPU or NVIDIA · Digital download
32GB RAM, CPU only? Start with a small model.
16GB RAM + 12–16GB NVIDIA VRAM? Start with 7B at 4K context.
ONE COMPLETE METHOD. TWO WAYS TO USE IT.
Both editions take you through the core build. The bundle saves file creation and adds tools for operating what you built.
STANDARD EDITION · V2.1
For readers happy to create their files from the guide’s full code listings.
Standard PDF · delivered through Gumroad
PREMIUM BUILDER BUNDLE · V2.1
For readers who want editable files and a repeatable way to inspect, measure and back up the stack.
Premium PDF + starter ZIP · delivered through Gumroad
A self-directed technical guide. You’ll use a terminal, need administrator access and download third-party software. Completion time depends on your hardware, downloads and starting point.
LOOK INSIDE VERSION 2.1
The guide separates “the API is reachable” from “the model actually generated a response.” Each stage has an observable finish line.
Offlinepython3 stack.py brief
python3 stack.py ingest
python3 stack.py ask "When does the backup run?"A briefing must create a file. Retrieval must identify its source. An agent must show the tool it actually called.
Shown: a page from the included v2.1 PDF. The workshop facts are labeled exercise data, not a customer success story.
A CLEAR FIT, BEFORE CHECKOUT
You can copy commands, follow a checkpoint and pause when something fails. You want a useful first stack, not a promise that one GPU replaces every cloud service.
Take the free fit checkUbuntu Server 24.04 on a dedicated x86-64 machine, or an existing TrueNAS 25.10 system. CPU inference works; NVIDIA acceleration has its own steps.
Native Windows, Apple Silicon, AMD GPU acceleration and DGX Spark installation are outside this edition. Those machines may run local AI; they need a different setup route.
Start with what you own. The hardware companion keeps upgrade research separate from getting your first model working.
BEFORE YOU BUY
Yes. It includes the full core method and source listings for the setup, briefing, retrieval and bounded-agent exercises. The bundle supplies editable copies and extra operational tools.
You’re buying an ordered route, explicit settings, complete exercises and checks that distinguish a running service from a useful result. Ollama, Open WebUI and their documentation remain available independently.
No. Installation downloads software and models. The exercises use local inference; external providers, search or connectors can send data outside your machine if you add them.
Premium-only: future version 2.xx revisions of the included guide and starter files, accessed through your original Gumroad purchase page when released. No automatic installation or release schedule is promised. Standard has no update entitlement.
The core build creates a small agent with one approved read-only tool. General-purpose frameworks are separate projects with their own model, context and integration requirements.
START WITH A WORKING BASELINE