Hello from Cairo Labs

A few weeks ago I was sitting at my desk at 1 a.m., watching a stack of AI agents build, test, and quietly reject each other’s work on a machine I’ve nicknamed Sparky. No human in the loop for that particular stretch — just models talking to models, shipping code. I’ve spent thirty years in IT, and I can tell you: that moment did not exist a couple years ago. It exists now, on hardware sitting three feet from my desk.

That’s what Cairo Labs is about, and this is where I’ll show the work.

Quick introduction first, because if you’re going to spend time reading what I write, you should know who’s writing it.

I’ve been a technical person my whole career. My proudest job was when I joined Aruba Networks in 2007 and spent over fifteen years architecting and managing enterprise network deliveries. Next I went to a network-as-a-service startup where I helped build out the products and the Professional Services offering — writing the pricing and scoping and structuring the consulting model from scratch. I watched AI rewrite the network engineer’s job from the inside, and decided to build with it rather than around it. I first formed Cairo Networks to deliver next-generation networking directly.

The constant across all of it: I like taking something complicated and making it work simply, reliably, and honestly.

These days the hardest and (to me) the most interesting problems are in AI, so that’s where I’ve pointed myself.

What runs here

Cairo Labs is the engineering side of that work, and it runs on real hardware — including our flagship, an NVIDIA DGX Spark AI workstation (Grace Blackwell GB10 chip, 128 GB RAM). On it I run local models, build practical tools, and operate a multi-tier system where AI agents plan, build, and QA software with me at the controls. It’s called Project Lumbergh, and it’s themed after Office Space. Sparky came online in November 2025, and the pipeline has been filing and building real work since soon thereafter.

Alongside it: things like a keyword tagger that runs a small model locally so a 200,000-image photo library becomes searchable without any of it leaving the machine, a tool that gets images onto a Samsung Frame TV without fighting Samsung’s requirements, a newsletter generated in-house on our own pipeline, and an ARM64 driver build I made for hardware the toolchain hadn’t caught up to yet.

Why in public

I’m not going to write press releases. I’m going to write build logs — what I made, how it actually works, and what I learned building it. Real numbers, real tool names, real commands. When something I built is useful to someone else, I’ll hand it over; a couple of these are staged for release under an MIT license for exactly that reason.

The AI space is full of confident people who’ve never shipped anything. I’d rather show you what’s actually on the bench and let you judge for yourself.

The rules

I have exactly some rules about AI, and you’ll see them in everything here:

• It’s powerful.
• It’s wrong a lot.
• Use multiple AIs to collaborate.
• Keep humans in the loop.

I’m an enthusiastic adopter — I’ll happily point three different models at a problem and let them argue. But I verify, I cross-check sources, and I don’t ship anything important on a single model’s say-so. Today’s tools don’t have to be perfect. They have to be better than the alternative, and you have to know where they fail. That’s the whole job.

Those rules are why the system is shaped the way it is. The model that writes the code is never the model that reviews it. High-stakes designs go past a panel of outside models before anything gets built. Anything consequential escalates to a person — and the interesting engineering is in deciding precisely which things those are, because a system that asks you about everything is just a slower version of doing it yourself.

What to expect

Notes from the workbench. How the pipeline is put together and what each piece earned its place by doing. Local-model infrastructure and the unglamorous plumbing that makes it reliable. ARM64 builds for hardware so new the software ecosystem hasn’t caught up yet. And the practical stuff I’d have wanted someone to tell me a year ago — which reviewer models are worth the tokens, what to put in your standards docs, where to draw the line between what the agents decide and what you do.

Welcome to Cairo Labs. Let’s build some things and be honest about how it goes.

The projects, experience and opinions here are mine. AI helped me turn my notes and build records into this piece and polished it for Cairoglyphics.ai.