Writing
Platform engineering, DevOps, and building with AI coding agents — the practical version.
A real WordPress + Elementor site rebuilt as a static Astro site: preserving five years of URLs, rewiring 1.7 GB of images that pointed at three different dead hosts, and the last step nobody talks about, which is destroying the old server.
What ChatGPT, Claude, and Gemini actually are, in plain English. Why they aren't plotting anything, why one of them knew about your family, and where the real risks actually live.
Most explanations of RAG stop at 'chat with your PDF.' For a long-running agent, retrieval isn't a one-shot lookup: it's how you decide, at every step, what belongs in a limited context window and what gets left out.
Every open port on a droplet is a standing bet against the entire internet. Here's how I stopped taking that bet: the real Tailscale setup behind this site's own analytics stack, the three things that broke before it worked, and the one port I still haven't closed.
Every project I own has its own CLAUDE.md, and none of them can override the one thing that actually matters. How a three-layer policy hierarchy turns AI governance into filesystem permissions instead of a promise.
Generic invoicing tools don't fit a solo consultant's actual pipeline, so I built Helm: a local-first back office where proposal, SOW, invoice, and tax set-aside all live in one connected model. This is the dogfooding story, and the decision that almost shipped it as the wrong kind of app.
GitOps promises that merging to main quietly becomes a live deploy. Here's the merge that looked perfect everywhere except in production, the exact error that explained it, and the boring fix that makes sure it can't happen twice.
Should an AI agent be allowed to edit its own guardrails? The answer is yes, if you separate the control plane from the data plane. Here's how, and why it's the safety.
The critics of AI-generated code are mostly right, and that's exactly why 'vibe coding' doesn't describe how I work. The whole line is: who reads the code?
A working engineer's ground-up map of how LLMs, tool use, and AI agents actually operate. No hype: just a stateless next-token function, wrapped in a loop.
I run two dozen GitHub repos as one estate with five small CLIs named for Odin's ravens and wolves. Here's the pack, and why the myth is really just good ops.
AI coding agents are brilliant, tireless, and eager to a fault, which is exactly what makes them dangerous. Why speed quietly defeats review, and how a handful of automated guardrails keep an over-eager agent from merging something regrettable.
No posts match your search or filter.