What I Learned After a Month of Building With AI

A month ago I published the first post on this blog. The idea: build a fully automated trading system with AI, without typing a single line of code myself, and document everything in public. I did not know if it would work.

Today there is a running system. It sits on a paper account, it pulls data every morning, calculates factor scores, generates signals and places orders. The live start with real money is planned for late June. Twenty-five blog posts later, this is what a month of building with AI actually taught me, including the parts that were not fun.

Speed was never the problem

The first surprise: AI builds individual pieces absurdly fast. A Python backend was up in a morning. Twenty-three thousand historical scores in one afternoon. A factor module that would take days by hand was done in fifteen minutes. And that is not a simple module. It weighs around twenty variables and categorizes them, with all the edge cases that come with that.

And still, the whole thing took a month. The system is now around thirty thousand lines of Python code. Individual pieces are done in minutes, but building them into one working production system, integrating, testing and debugging, that took four weeks.

Because speed only moves the bottleneck. Typing code is no longer the slow part. The slow part is deciding what should be typed. Which architecture is right. Whether the output is actually correct or only looks logical. An AI that is wrong is wrong just as fast as it is right.

What was harder than expected

I underestimated three things.

The first: data is never uniform. Walmart has a fiscal year that does not line up with the calendar, and that broke my earnings ingest completely. Hundreds of companies do this. Every assumption about how data behaves dies on an exception.

The second: paper trading is not the same as live. Some bugs only show up in production, because that is where you finally have the real volumes and the sharp timing. A sizing bug set every new BUY to zero shares. It was logically wrong, but only visible with seventy-six candidates at once.

The third: it runs does not mean it works. My momentum calculation ran without errors and the top twenty looked logical. Done, I thought. A few hours later I found the mistake. Running cleanly and being correct are two different things.

What I would do again

Record everything in files. Decisions, lessons, content moments, the working agreements with the AI. Not in my head, not in a chat that will soon be gone. In files in the repo. It took discipline, but it is the reason I still know exactly why I made certain choices four weeks later.

And building in public. Writing a blog post forces you to understand what you did. A few times I only spotted a flaw in my thinking while writing the explanation.

What I would do differently

Trust how it looks less. The most expensive mistake of the month was a setting I had made permanent and never removed. The result: for two weeks the system processed stale data without a single error message. No error, just quietly wrong. Since then I am more suspicious of anything that works without complaining.

And fewer marathon days. Several sessions ran long, and it was exactly in the late hours that I made the sloppy choices. For a system that will soon trade with real money, pushing on while tired is not a virtue.

The takeaway

AI removed the typing, not the thinking. That is the whole lesson in one sentence. I no longer write code, but I still have to understand the architecture, check the output and catch the subtle mistakes. The discipline a production system demands has not shrunk. It has only shifted to where it belongs: with the human who is responsible.

In a month I will know whether the system can also make money. That is a different question. But building, that works. I am sure of that now.