One person, a handful of products.

nuumoo is a handful of things I built because they interested me, or because I needed them myself.

Two products, WatchMyCalendar and echo. Three websites, including this one. The infrastructure they all run on. And a few tools that so far have only been useful to me.

Nobody asked for any of it. I didn't have to build it. I wanted to.

And almost all of it was built with AI.

How I work with it.

Most of the work happens in Claude Code. ChatGPT sometimes, for a second opinion or a quick question on the side.

The pattern is mostly the same. I describe what I want, why, and how it should fit into what's already there. The AI proposes and writes. I review, push back, change direction — and decide what goes in.

The useful part isn't that the AI writes code. It's that I can build things I would never have found the time for otherwise.

Where it carries.

Breadth.WatchMyCalendar is Kotlin, echo is Python, the websites are Astro, the infrastructure is Docker stacks kept as code. One person covering all of that, at this pace, wasn't realistic a few years ago.

The tedious parts.Boilerplate, migrations, tests, the fifth variant of the same form.

Getting started.An empty folder is a lot less daunting when something can sketch a first version in minutes.

Where it doesn't.

Deciding what to build.The AI doesn't know which problem is worth solving, or which feature should be left out.

Architecture.It will happily build whatever you ask for — including the wrong thing, very convincingly. So the architecture is mine: what the pieces are, how they talk to each other, what runs where. The AI is a useful sparring partner in that discussion. The decision isn't its to make.

Review.Output that looks right isn't always right. Twenty years of reading code help more than ever here.

Operations.When something breaks, it's my server, my backups, my responsibility.

What about local models?

Running everything locally would fit nuumoo well — no cloud, no data leaving the house. So I tried: Ollama, Open WebUI, and llama.cpp, which is faster but a lot more work to set up, with CUDA and everything around it.

On the hardware I have, the results weren't good enough for daily work. Too slow, or too far behind the large models when it comes to code.

Images are a different story. The pictures on this site were generated locally with ComfyUI, and that works well. A good reminder that where AI fits depends a lot on the task.

What changes.

I still make the decisions. I still read the code. I still run it in production.

What changed is how much one person can build — if they know what they want, and can tell when the result is wrong.

That's probably the most useful skill right now. Not writing every line yourself, but knowing which lines matter.