On this same stage I have built an arcade game in 90 minutes and a Micro SaaS in an afternoon, live, with AI writing most of the code. Those talks end with applause and a working demo.

This one starts the day I handed the keys to somebody else.

I have shipped a few of these for real: AI Radar and Momentum, both MVPs, plus a multilingual AI podcast pipeline that actually publishes on a schedule. Not a million users, nothing dramatic. A handful of beta testers and a product that has to keep working when I am not watching.

That was enough.

Because here is the thing nobody says: you do not need scale to find out what the AI got wrong. Five people poking at your MVP will surface the permissive security defaults the model quietly chose for you, the error handling that only covers the path you described in the prompt, the dependency nobody ever audited, and that special moment when you open your own repository three weeks later and cannot remember how any of it works.

We'll go through what actually broke, what I check now before anything gets a public URL, and where human attention genuinely pays off, which is the data model and the trust boundaries and almost never the syntax. Then we'll look at the same pattern one floor up, because as a fractional CTO I watch engineering teams ship AI generated code to far more users than mine and repeat exactly the same mistakes at a scale where they cost real money.

Vibe coding is real and it is here to stay. This is what it costs you the moment somebody else depends on it.

You do not need a thousand users to find out your AI wrote bad code. Five are enough! I shipped AI Radar, Momentum and a multilingual podcast pipeline as MVPs, handed them to beta testers and watched what broke first. Permissive security defaults, error handling that only covers the prompt, dependencies nobody audited, and code I could not read three weeks later. Plus the checklist I now run before anything gets a public URL!