Adventures and experiments with AI
Natural language, pseudocode, and code are not competing tools. They are different languages for different levels of precision.
In the age of AI-assisted programming, learning to program is not only about writing every line of code by hand. It is about knowing when an idea needs to become more precise, how to control that translation, and how to distinguish between decisions made by a human and decisions filled in by an AI system.
Proof-of-concept research. github.com/amiune/theblackboxapproach
AI is great at generating code. It's also great at rewriting your entire project when you asked it to tweak one button. BlackBoxApproach is our experiment in fixing that.
Proof-of-concept research. github.com/amiune/MokuML
What if a small business could connect Postgres, describe a prediction in plain English, and get results on their Mac without a data science team, cloud ML stack, or Kumo-style subscription?
That's what we're exploring with MokuML. Think lightweight Kumo.ai: connect a database, ask for predictions, get answers. But everything runs locally.
2025 has been a year of hands-on experimentation for us with LLMs and applied AI. Instead of just watching the space from the sidelines, we spent the year building, testing, breaking things, and trying ideas that sometimes worked and sometimes didn’t.
The overall feeling is mixed. Some parts of the tech are impressive and moving incredibly fast. Other parts still feel early and fragile. Here’s a summary of what we learned and how it’s shaping what we want to do next.
TLDR: We worked on implementing LLMs, Vector DB, and RAG for elMejorTrato.com. We concluded that:
To learn Artificial Intelligence from scratch, FreeCodingTour offers three courses that should be taken in the following order: