Local AI Setup Kit

The situation you are in

You have material you cannot paste into a chat box. Maybe a contract says so. Maybe your employer blocks it at the proxy. Maybe it is a client file, an investigation, a salary band, a draft you are not ready to hand to a vendor, or notes about a person who did not consent to being processed by anybody. So you do the work the slow way, while watching everyone else get an hour back every day.

Or you have no constraint at all and you simply want the thing to keep working — the same model, the same behavior, no per-token bill, no rate limit at four in the afternoon, no email announcing that the version your workflow depends on is deprecated in ninety days.

Either way you have heard you can run one of these on your own computer, and every guide you have found either assumes you already know what quantization is or is a fifteen-minute video where someone with a very expensive desktop says it is easy. You do not know whether your laptop can do this. You do not know which of the four hundred model files on the download page to pick. You suspect that if you spend a Saturday on it you will end up with something that answers slowly and gets things wrong, and you will have learned nothing about whether that was your hardware, your model, or your prompt.

What this kit is

The afternoon, written down. Nine chapters and eleven files that take you from "I have a laptop" to "there is a private assistant on it that does one job for me every day."

It is opinionated where opinions save you time. It says memory is the only hardware number that matters and tells you to buy memory over every other upgrade. It says start at four-bit quantization and take the bigger model, not the higher precision. It says the system prompt is worth more than the next model size up. It says paste the whole document before you build a retrieval pipeline. It says run one job for two weeks and delete it if it does not earn its keep. Where reasonable people differ, the templates leave room — but the kit takes a position rather than listing options, because a list of options is what you already have.

It is also honest about the ceiling. Chapter 1 states plainly what a model that fits in your memory cannot do, because the fastest way to abandon local AI is to expect it to replace a frontier model and discover in week two that it does not.

What is inside

• The guide. Nine chapters: why local and the honest ceiling; hardware, where memory is the number; models, sizes, and quantization in plain words; the install afternoon on all three platforms; the three settings that matter; grounding it in your own documents; what stays local and what goes to the cloud; one daily job wired to run; and keeping it running a year later. About 14,500 words.
• Sizing and picking. A hardware sizing record you fill in for your own machine, carrying the table from 8 GB to 64 GB with speed ranges, the headroom rule, and an upgrade-request section you cannot fill in on vibes. A model-pick worksheet: five questions from your memory number to one specific download, plus the twenty-minute head-to-head that actually decides it.
• Install. Separate runbooks for macOS, Windows, and Linux, each with the failures that happen on that platform — the PATH that did not refresh, the download a proxy stalls, the graphics card the runtime cannot see, the antivirus that quarantines the binary — and each ending in the record of what was installed.
• Running it well. A settings card for temperature, context length, and system prompts, with three worked cards you can steal. A document-grounding setup covering corpus discipline, the refusal instruction, the citation rule, and the ten-question acceptance test that tells you whether the answers can be trusted.
• The rule. A four-tier routing rule and a six-step redaction path, both written to be marked up by your security reviewer rather than adopted silently, with sign-off blocks.
• The daily job. Daily-job scripts in two versions — a fifteen-line shell script and an HTTP API version that returns JSON — plus the prompt files, the scheduling notes for each platform, and the two-week tally with a kill criterion you commit to in advance.
• Maintenance. A troubleshooting checklist with the diagnostic order, thirteen symptoms and what each one actually is, a twenty-minute monthly list, and the yearly hour.
• START-HERE. One page that tells you what to read depending on whether you have an hour, an afternoon, a day, or a week, and which file to open at which moment.

Every template uses [BRACKETED PLACEHOLDERS] for your specifics and carries a "How to adapt" note. They are plain markdown, so they paste into a word processor, a wiki, or a ticket without fighting formatting.

Who wrote it and why he can

Wayne Bridges, a public-sector solutions engineer who builds and runs AI agent systems daily. The guidance about what these models do, where they break, and how much memory it actually takes comes from running them, not from reading about them. The chapters on the ceiling and on grounding exist because both lessons were learned the expensive way.

He is not an attorney. The chapters that touch licensing, confidentiality obligations, or organizational policy say so at the top and are built to be reviewed by the person who owns that risk.

What this is not

• Not a benchmark roundup. No leaderboards, no numbers that cannot be supported. Speeds are ranges. The kit teaches you to run your own ten prompts against two models in twenty minutes, which beats any comparison written by someone with a different machine and a different job.
• Not a promise that local replaces the cloud. It is the opposite. The routing rule is a core deliverable precisely because the two-tier habit is what makes this survive past month one.
• Not a fine-tuning or model-training course. You will download weights and prompt them well. Training is a different project with a different budget.
• Not legal or compliance advice. The routing rule is a draft for your counsel and your security team, not a substitute for them.
• Not a permanent list of the best models. Model names move monthly. Chapter 3 teaches you to read any model name and pick from your own memory number, which does not expire.

Who should not buy this

• Anyone with an 8 GB machine who wants frontier-quality answers. Chapter 2 will tell you honestly what you get, and it is not that. Better to know now.
• Anyone who will not open a terminal. There are graphical front ends and the kit covers one, but the install, the scripts, and the troubleshooting all assume you can paste a command and read what comes back.
• Anyone looking for a zero-maintenance appliance. This is twenty minutes a month, forever.
• Anyone who needs this to satisfy a specific regulatory obligation. That conversation starts with the person who owns the obligation. Bring them chapter 7; do not bring them a download.

If you have a machine with 16 GB or more, work that involves text, and a reason — a rule, a contract, or just a preference — that the work should not leave your desk, this is for you. Read the sizing table, pick your model, and give it an afternoon.