Notes · 9 posts

Notes

Opinions and field notes on software, AI and building things on my own time. Written when I have something to say, not on a schedule.

October 2026

  1. How I used AI to control my Bluetti stationsOne of my Bluetti power stations has no Wi-Fi. This is how I used AI coding agents to turn a PC at home and an iPhone app into a way to see and control it from anywhere, with an AC output that turns on only when the power goes out.iosreact-nativebluetoothhome-automationai-agents6 min readenpt
  2. Sonnet beat Opus on my security evals for under half the costI added model evals to my quality gate: contenders, a price table in US$ per 1M tokens, prompt and code datasets and three security sets. Sonnet 5.5 passed 30 of 30 medium tasks where Opus 5.5 passed 27, and one of my own rubrics turned out to be wrong.quality-gateevalsllmsecurity8 min readenpt
  3. I stopped pasting API keys into my AI agent's chatEvery key you paste into an AI agent's chat goes back to the model with each message after it. I built a small plugin that hands the agent a key without the conversation ever seeing it.claude-codepluginsecretsredaction4 min readenpt
  4. I gave my webcam a lighting crewA video call in a dim room, a flat grey face, and one question: what if the lamps were software? The story of Virtual Lighting, an app that lights your real room, and the surprises along the way.webgpuelectronchrome-extensioncomputer-vision3 min readenpt
  5. My quality gate got a mission controlI rebuilt the Quality Gate dashboard around one idea: every answer should be a click away from where I'm looking. Side panels for anything you click, a run page that assembles live, a Ctrl/⌘+K palette, and a background that knows when a run is going.ai-agentsqualitydashboarddesign5 min readenpt

September 2026

  1. I built a quality gate for AI deliveries. Here's how I specified itA walkthrough of the quality gate I specified for AI deliveries, from ticket and diff to checks, judge, scorecard and calibration.ai-agents quality evaluation testing5 min readenpt
  2. Marketing Effect: The AGIThe more I use and compare these models, the harder it is for me to separate real progress from hype and marketing. At this point, I care less about AGI claims and more about whether AI can actually deliver reliable results without burning through time, tokens, and money.agiclaudeopenaianthropicroireturn on investment3 min readenpt
  3. How this site is builtStatic Angular, no server, a strict CSP and a CI gate that measures bytes instead of scores. What is in the repo, what each piece is for, and the two decisions I would take back.angularperformancecloudflare7 min readenpt