Michael Limberger
Need me? Email mike@limberger.ca
AI
The Ai Watchman - Overview
AI watching AI
This talk is a local image-review pipeline. A vision model scores pictures on hardware you control, so the files do not need to go to a hosted service.
What if AI could watch AI?
That is what we are exploring tonight. We will use a vision model (an AI that can see and understand images) to automatically check the output of an image generator. The whole pipeline runs on your own machine. No cloud services. No Big Tech looking at your pictures.
Hundreds of pictures, one watcher
When you generate hundreds of AI images, you cannot manually review them all. You need automation. But how do you automate "does this look okay?"
The answer: use another AI to look at the images and tell you.
| Step | What happens |
|---|---|
| 1 | AI-generated images |
| 2 | Vision model analysis |
| 3 | Pass or reject |
The walk tonight
By the end you will understand what a vision model is and how it "sees" images, how to send images to a local AI for analysis, how to get structured, parseable responses, and how to use pattern matching to make pass or fail decisions.
We do everything the raw way using basic command-line tools and bash scripting. No special scripts to install. Everything here you can type yourself and adapt.
What we are not doing
We are not giving you a finished tool to download and run. The goal is the fundamentals so you can build your own automation.
By the end you will have working curl commands you can copy, bash scripts for batch processing, and the knowledge to adapt this to your own needs.
What you need in the room
A computer (Mac, Linux, or Windows with WSL). Terminal access. Willingness to type commands. That is it. We will install everything else as we go.
An LLM is a large language model: a program trained on a huge pile of text so it can continue language. A vision model is an LLM that was also trained on pictures.