The Workbench collects an academic’s workshop notes on two intertwined lines of work: building small, transparent tools for teaching finance, and running hands-on experiments to understand how AI language models behave.
The writing comes in two registers. Long essays work an idea through carefully — the kind of piece meant to be read once, slowly, and returned to. Field notes are shorter and rougher: a result, a dead end, a tool that just started working.
The aim throughout is the same: to make the machinery legible. Whether the subject is a spreadsheet model or the inner workings of a Transformer, the bet is that you understand a tool best when you can see — and ideally perturb — every moving part.
I am Luca Erzegovesi, and I write these notes from my own teaching and research.
A quiet site, on purpose
These are static pages on GitHub Pages. No ads, no cookie banner, no popup asking for your address, no paywall, no algorithm choosing what you see next — and no cookies set by this site. People find their way here because someone sent them a link, and that is the door I would like them to keep coming through.
How this is made
The work here is mine, but not made alone. I draft and edit the essays — and build and debug the small tools and models I demonstrate — in close collaboration with Claude, Anthropic’s AI assistant, often in extended working sessions. Claude helps with drafting, sharpening explanations, writing and debugging code, and stress-testing arguments; I set the direction, make the judgement calls, verify the results, and take full responsibility for what is published. Where a piece leans especially heavily on that collaboration, I say so in the post. AI is a named tool in this workshop — not an uncredited ghostwriter, and not a co-author.