Blog

What it actually takes to run AI agents in production — infrastructure patterns, security decisions, and lessons from someone who did the same for cloud a decade ago.

  1. Publish Your Personal Website on GitHub Pages, for Free, Using Claude Code

    A step-by-step walkthrough from a fresh Mac to a live website at yourusername.github.io. One page of plain-English spec; Claude Code handles the build, the repo, the deploy, and the optional custom domain. About 45 minutes including the one-time install.

    · 10 min read tutorial claude-code github-pages static-site getting-started

  2. Confirm It's Still Fast

    A leftover extra_hosts entry in a generated docker-compose.yml silently routed every local-LLM request through the host gateway for an unknown period. Nothing alerted. A ~200 LOC regression harness is the only thing that would have told me — and the only thing that will tell you when yours breaks next.

    · 6 min read local-inference ollama performance agent-ops openclaw

  3. The Fix Was the Speedup

    A drifted systemd override left an unauthenticated LLM endpoint on a public IPv6 and pinned gemma4 inference to an Ollama binary that hadn't yet flipped flash attention on. One cleanup closed both — same hardware, same weights, same vendored llama.cpp.

    · 9 min read agent-ops local-inference security infrastructure day-2-operations

  4. Seven Bugs, Zero Model Problems

    After a ClawHQ rebuild, Clawdius was completely unresponsive. Seven distinct bugs. Every one was a platform failure — config drift, firewall chain reactions, GPU defaults that didn't match real workloads. The model was fine the whole time.

    · 3 min read agent-ops debugging local-inference deployment day-2-operations

  5. Don't Blame Your Layers

    When a customized agent deployment breaks, the instinct is to simplify. We almost ripped out our security stack to fix an upstream bug. The update pipeline saved us from ourselves.

    · 3 min read agent-ops debugging infrastructure day-2-operations