On the Thermodynamics of Espresso
and Other Unsolved Problems

Gabriele Salvo
Department of Caffeine-Driven Engineering
Eindhoven, The Netherlands
February 2026 (preprint — not yet peer-reviewed by my mother)

Abstract

We present a longitudinal study on one Italian specimen relocated to the Netherlands, examining the correlations between espresso intake, lines of code produced, and the probability of explaining AI to someone who just wants a chatbot. The author has taught 150+ students, shipped 25+ projects, and won awards for building bionic limbs and haptic devices — none of which impress his barista. A live multiplayer experiment is included below.

1. Introduction

Let \(G\) be a Gabriele. Let \(E(t)\) denote the number of espressos consumed by time \(t\). We observe empirically that:

\[ \texttt{productivity}(G) \propto E(t)^{2} \cdot \left(1 - P(\text{meetings})\right) \]

Born in Italy, currently based in Eindhoven, the author divides his time between teaching machines to think and teaching humans to think about machines. This site exists because every Italian needs at least one side project that has nothing to do with making money.

2. Experimental Setup (What I Actually Do)

By day, I help companies build AI without accidentally breaking EU law. By night, I build things that break in different, more creative ways.

Definition 1 (AI Engineer). A person who spends 80% of their time cleaning data, 15% explaining why the model needs more data, and 5% doing the cool stuff from the conference talks.

The instruments used in this ongoing experiment include:

Theorem 1 (The Certification Conjecture). If you have enough certificates, at least one person at the meeting will take you seriously.
The author holds 8+ certifications including Certiport IT Specialist in AI and Python, and is a certified instructor. The conjecture remains unproven in the general case. \(\square\)

3. Notable Results

Against all odds and despite a persistent espresso dependency, some things actually worked:

3.1. Bionic Prosthetics (PhD+ 2024 Winner)

Built predictive AI for prosthetic limb control — making a mechanical arm anticipate what you want to do before you know yourself. Target latency: \(\approx 0\) ms. Your arm, but with autocomplete.

3.2. Immersive Haptics for Healthcare

Designed a modular haptics suite for bedridden patients using HaptonTech polymer. Won the ICMS × HaptonTech Master Challenge 2025. 35% cheaper than comparable kits. The patients liked it. The spreadsheet people really liked it.

3.3. Teaching (The Infinite Loop)

150+ students across Save the Children, CoderDojo Pisa, Hack Your Future, and Volta Institute. Completion rate: 90%+. The remaining 10% are presumably still debugging their first for loop.

Lemma 1. Teaching someone Python is easy. Teaching them that pip install and conda install should never be mixed is the real challenge.

4. The Experiment (Your Turn)

To demonstrate real-time state persistence and prove this site isn't just a static PDF cosplaying as a webpage, every visitor shares a single chessboard. Moves are stored in Redis and synchronised globally. Chaos is guaranteed.

♟ Play Chess Against the World

Current hypothesis: a chessboard played by the entire internet converges to checkmate in finite time. Early results are... inconclusive.

5. Open Questions

6. Conclusion

The results suggest that transplanting an Italian to the Netherlands produces a non-trivial amount of AI engineering, a deep appreciation for heated bike paths, and the unshakeable conviction that no meeting should start before the second espresso.

For the professional version of this, see salvoaistrategy.eu. For everything else, you're already in the right place.

Q.E.D.