A few of the pieces most relevant to the work are below. The full archive is on Medium.
I started vibe-coding the moment the tools were good enough — and inherited a new problem: I'd handed an AI assistant the keys to my AWS account, my repos, and my local drive. The containerized setup I built so I could let it run without losing sleep.
Two major rewrites: one became a thriving company, the other delivered technically but missed the business return. A spreadsheet model for the decision — where the value isn't the model's accuracy, it's the discipline of building it.
The first IoT cycle, I was at the helm but really a passenger. The second time, I brought software's tools to the hardware team — weekly iterations, and a first over-the-air firmware update that didn't brick the fleet. What carried over from software to hardware, and what didn't.
You can't test a non-deterministic app with deterministic tests — we learned that running a production RAG chatbot. The behavioral-testing approach we built, and why anyone shipping LLM features needs one.
Every consulting firm eventually hears the siren call: let's build a product. It almost never works. I've run that transition twice — and the reasons it fails aren't strategy or market, they're cultural and structural, and always underestimated.
A full production chatbot architecture that runs at about $10 per 1,000 conversations. The decisions, the trade-offs, what to ignore from the vendor noise, and the patterns worth stealing.
Technical debt isn't a problem to eliminate; it's a tool to use. Like financial debt, the question isn't whether to take it on — it's whether the return beats the interest. When to incur it, when to refinance, when to pay it down.
More data doesn't make a better chatbot — better curation does. A framework for populating vector databases from messy real-world sources, and tuning what actually goes in.
Not every company needs a full-time CTO — some need one for a season, some for a single hard problem, some just often enough to avoid expensive mistakes. When the fractional model fits, and when it doesn't.