What this is and why it exists

The gap between a model that passes a notebook cell and one that runs at 2am without breaking is wider than most people are told. Closing that gap requires a specific kind of thinking — systems design, operational discipline, and a willingness to engage with the unglamorous parts of production ML. That’s what this blog is about.

I’m Carlos. My background includes the United Nations, IDB, Yale, Mercado Libre, and Globant — plus a few years teaching ML across Latin America and Spain. The through-line is the same problem in different institutional forms: making machine learning systems that actually work, for the people who need to use them.

I write here when I have something specific enough to be useful. Topics cluster around three technical pillars — LLMOps, AI software engineering, and GCP — and one parallel research thread that I find equally interesting: applying NLP and LLMs to music analysis.


Reading paths

LLMOps — MLOps for the LLM era

The MLOps discipline that emerged around 2020 was designed for classical ML: feature pipelines, model registries, batch inference, drift detection on tabular data. LLMs broke most of those assumptions. This is where I work through what the new discipline looks like.

AI Software Engineering

Training a model is 20% of the work. The other 80% is the system around it. These posts cover that layer.

Music & AI

What happens when you apply NLP, embeddings, and LLMs to music? I’ve been exploring this seriously — with statistical rigor and genuine curiosity about where the tools fail.

Full Music & AI section


How to follow

The best formats are RSS and the newsletter. I write when I have something worth saying — not on a fixed schedule.

GitHub · LinkedIn · X