← All series Live · 7 episodes

HurricaneWise: a small product, run the way a company would run one

A hurricane tracker taken from a weekend idea to real production by one person plus AI, then handed the job of repairing itself. Seven episodes, receipts included.

The point: The tracker is the vehicle. The question underneath it is whether a product can be built with AI, built properly, and then run without a person in the loop. Operations answer yes today. Product judgment is still mine, and that gap is the interesting part.

Start with the argument → The whole case in one read: the expensive part was never building it. These episodes are its deep dives.
  1. 01 Three weeks, 184 commits, and 91 of them were fixes Half of everything I wrote in the first three weeks of HurricaneWise was me correcting something that already looked finished. That ratio is the honest headline of building with AI, and it is the reason the rest of this series is about operations rather than code.
  2. 02 The phantom storm, and four other bugs that rebuilt my monitoring Five real failures from the first months of HurricaneWise, each with the rule it left behind. They are the reason I trust the dashboards now, and four of the five were invisible from inside the system while they were happening.
  3. 03 Seven services and one rule about where work lives Choosing the services took an afternoon. Deciding which work belonged on which of them is the decision that kept the bill near zero, kept the pipeline punctual, and made the security model a property of the architecture rather than a feature bolted on at the end.
  4. 04 Nine specialists, one contract each There is no prompt in this system that "does the AI." There are nine, each with one job, one required output shape, and a set of honesty rules I wrote after catching it being wrong. Swap the model and the behaviour holds, because the behaviour lives in the instructions.
  5. 05 $15 a month, itemized by the thing that spends it Building software got cheap and everybody noticed. Running it was always the expensive half, and that is the half this episode is about. About $9 a month became about $15, every dollar of the increase is a decision I can point to, and the numbers come from the system’s own machine-generated snapshot rather than from my memory.
  6. 06 The AI opens the pull request. The checks decide. When its own monitoring detects a failure, HurricaneWise goes from detection to a verified production deploy with nobody in the path. There is no approved list of failures it is allowed to touch: detection is the permit. What holds it is an envelope, a set of hard limits on how it may act, and a path that escalates to me the moment it cannot finish.
  7. 07 The job moved up a level Building got cheap and everyone noticed. Running the thing was always the expensive half, and that is the half now falling. This is what that did to my job, and what I think it does to who gets to build at all.