← All series Live · 7 episodes

HurricaneWise: one person running a hurricane tracker to a company's standard

I built a hurricane tracker with AI, took it to production, and then set it up to repair its own detected failures. Seven episodes, with the numbers.

The point: I wanted to see if I could build a product with AI to the standard a company would hold it to, and then stop doing the operations work myself. At this point repairs go out without me stepping in. What gets built next, and what to do with user feedback, is still up to me.

Start with the argument → Seven episodes on running it, with the numbers.

1. 01 The first three weeks of HurricaneWise The first working version of HurricaneWise took me about three weeks and 184 commits, and 91 of those were fixes to things I thought were done. After launch most of my time went into running it.
2. 02 The phantom storm and four other bugs Five bugs from the first months of HurricaneWise that changed how I monitor it. Four of them set off no alarm inside the system while they were happening.
3. 03 How HurricaneWise is put together The pipeline that runs every minute is on a $7.60 Hetzner box, the public site is on Vercel, and my estimate for the whole bill is about $15 a month.
4. 04 The nine prompts behind HurricaneWise HurricaneWise uses nine prompts and each one does one job. Five of them return data that gets checked against a schema, the post-mortem writer returns prose, and the resolver prompt drives a coding session. Most of the rules in them went in after I caught the output getting something wrong, and they live in versioned prompt files.
5. 05 Costs: about $15 a month HurricaneWise used to cost about $9 a month. Now it's about $15, and about $40 in a stormy peak month, mostly because the server got $2.35 more expensive and the AI runs repairs now too. The monthly numbers are estimates from my cost model, and the usage behind them comes from a script I run against every provider.
6. 06 How HurricaneWise ships its own repairs HurricaneWise ships its own repairs now. It attempts every failure its own monitoring detects. There is no approved list. The only exclusion is an incident in the resolver itself, and product decisions are still mine. When an attempt fails I get it with the evidence attached.
7. 07 What my job looks like now With AI writing and repairing most of the code on my hurricane tracker, my part is the rules, the budgets, the off switches and deciding what gets built.