Skip to main content
Approach

The gates around the work

AI writes most of my code now. It doesn't get a free run. The models live on my own hardware behind a set of gates, and every session runs inside checks that decide what the AI may read, what it may claim, and what may leave this machine.

Every way in

Requests to my local models pass through these gates. A request that fails one stops there.

Every gate is written out below.
  1. 1

    Known devices only

    A Tailscale policy lets in only devices I've tagged, and only to two doors. Anything else on the private network sees nothing.

  2. 2

    Key-only logins

    SSH logins use keys.

  3. 3

    One way to the models

    The model door forwards to a lock, and only while the lock is running. If the lock stops, nothing answers.

  4. 4

    The lock

    nginx checks a key for every request. No key, or the wrong key, gets a 401. It fails closed.

  5. 5

    Every call counted

    A gateway records which caller used how much of the Ollama models, and it only listens inside the machine.

  6. 6

    Models that stay inside

    Ollama only listens inside the machine. vLLM and CopernicIQ, my decision service, are reachable only through an SSH session.

How a key is made

A new key is generated in memory, stored in Azure Key Vault and Windows Credential Manager, and handed to the machine through SSH's input. It never appears on a command line, in a log or on screen. It only goes live after a check that needs administrator rights.

How code gets there

Only reviewed code, packed from a named commit, with a hash that's checked before anything runs. It's installed read-only for everyone but the administrator, and the installers refuse to run from anywhere else.

Inside a session

The same rules load into Claude Code, Codex, Grok and Antigravity from one source I review like code, and GitHub Copilot is next. Inside a Claude Code session, checks run at every step.

Every check is written out below.
  1. 1

    Before anything is read

    A sensitive-data scan, built on Microsoft Presidio, looks at the prompt and at each file or command before it runs. It records what kinds of data it saw and how many, never the data itself.

  2. 2

    After every tool

    A fabrication guard flags identifiers the session never actually saw, and an encoding guard catches damaged text.

  3. 3

    Before an answer stands

    A quote check confirms that every file the answer names was opened and every quote was really seen. Quotes it can't match go to CopernicIQ for a second look.

  4. 4

    Before anything leaves this machine

    A filter applies each client's data policy. Anything shaped like a secret is held back. Network shares are never sent. Mixed folders go through flagged, and still only to my own hardware. What passes goes only to CopernicIQ on my own hardware, and only counts are logged.

Client work has its own seat

Client work runs in its own seat, with its own rules and security settings.

Built Something with AI?

I'll spend 30 minutes telling you what my production-readiness review would flag.