A study track toward a data-analysis agent
ArchivedTwo legs of a seven-week route are done — Python with LLM APIs, then Pandas for loading, cleaning and grouping. It stopped at SQL, and it says so.
This is not a project. It is a study track. It sits here because it produced something, and because it reached a conclusion.
The goal was never to write a few analysis scripts. It was to end up with a data-analysis agent I would actually use. The route: Python and LLM APIs → Pandas → SQL → analysis in practice → tool calling and agent architecture.
What I finished
- Python, and getting an LLM API to answer for the first time — virtual environments, the terminal, reading the key out of an environment variable, and a first reply on the screen.
- The core of Pandas — loading, cleaning, grouping. The closing exercise was a set of coffee orders: sales and average price per product, with the cleaned data and the summary table each written to disk.
Where it stopped
SQL got started and then stopped. The drive is gone for now, and I am not going to pretend otherwise.
Why it is still here
Because none of it was wasted, and because the direction has not changed — in the voice platform I built afterwards, the questions of “which agent does this go to” and “how does a tool get called safely” are exactly what the fourth leg of this route runs into. I went around in a circle and came back to the same place.
I will pick this one back up.