Marco
Serenelli.

AI engineer in Rome. Most days I'm wiring LLMs into production at Iamones, building retrieval pipelines, agents, and the eval loops that keep them honest. The interesting part isn't the model itself, but everything around it that decides whether the thing actually works for a real user on a Tuesday afternoon.

I learned that the hard way. A double Master's across Reykjavik and Camerino ended in a thesis on detecting stress from physiological data, months of cleaning noisy sensor readings before any model could touch them. That's where I learned a clean experiment is worth more than a clever model, and it shapes how I think about LLM systems now: the hard part isn't the model, it's everything upstream and downstream of it.

Reykjavik was also where NLP got its hooks in. My first real project was something small and classic, classifying emails as spam or not, probably with a BERT, I honestly don't remember. What I do remember is watching a classifier come alive on a real dataset and thinking, oh, this is crazy. I've been chasing that feeling ever since, through agents, RAG, and the whole orchestration layer that turns a model into something that actually does useful work.

A decent chunk of every week goes to reading. Papers that take two passes to click, GitHub repos I fall into for an hour, notes from people building similar things from different angles. The field moves fast enough that last year's intuition is usually outdated this year, so staying curious isn't optional.

Most papers don't survive contact with a real dataset, and the ones that do rarely survive in the shape the authors intended. That's the fun part. You read something on Saturday, hack a rough version by midweek, and by Friday you know which 10% of the idea actually holds up. Most of what I end up on Substack writing about is that 10%, me working through something until I figure out what I actually think about it.

I care about systems that ship, code reviews that teach, and the small decisions that nobody notices when they're right. When to let the LLM decide and when to hard-code the path, what you name the thing, where the prompt ends and the code begins. If any of that sounds like the work you're doing, the email's in the colophon.