Skip to the projects
¿D?

hi, I'm

Diego Ruiz Piqueras

Every project starts with a question, and almost all of them end with a more interesting one.

Software engineer · Artificial intelligence · Madrid

↓ ask me something

tap the photo: this is how my neural network sees me

the thread

Four questions that brought me here

Each one is a project, and each one has something you can play with.

2022 · Emporium Bot · Node.js

Can a program have a personality?

My first real project, from my early university years: a Discord bot with a mind of its own. It speaks Spanish and English, moderates users, does maths and, if you say hi, sometimes answers however it pleases. More than twenty commands and a lot of hours reading documentation.

What stuck with me: a program is designed for the people who use it, not just to work.

  • JavaScript
  • Node.js
  • discord.js

Code ↗

# general

    ↓ try a command

    …and then I wondered: what if it could also learn?

    2023 → now · LegofyNet · Generative AI

    Can a neural network learn to draw in Lego?

    A Pix2Pix network that turns a photo into a Star Wars Lego minifigure. It started as a team assignment at UPM and I've kept growing it: my own dataset of 2,328 images, scraped and cleaned by hand, 22 character classes and a website where the model runs right in your browser.

    What stuck with me: good data matters as much as the model, and the network's mistakes teach more than its hits.

    • Python
    • TensorFlow
    • Pix2Pix
    • TensorFlow.js

    Try it with your photo ↗ Code ↗

    Input photo
    what it gets
    Generated Lego minifigure
    what it draws

    real model outputs (64×128 px) ·

    …and then I wondered: if a network learns from what we show it, what prejudices does it pick up too?

    2025 · EQUITIA · Bachelor's thesis

    Does an AI answer differently if you change a single word?

    My bachelor's thesis, done with Telefónica in their TUTORÍA programme. It's a tool that asks a language model hundreds of questions, changing only who they're about (their age, their origin…), and measures whether the answers change with it. Six kinds of test, from closed questions to prompt injection attempts, and a report that flags anything out of the ordinary. I got a 10 out of 10.

    What stuck with me: measuring well is harder than building, and an AI can sound neutral without being neutral.

    • Python
    • LLMs
    • Hugging Face
    • Statistics

    Thesis ↗ Code ↗

    The figures are the real answers of the evaluated model (DeepSeek-R1-Distill-Qwen-7B). Another model wrote the questions from templates; here they are summarised. Without bias, every bar would be the same length.

    …and then I wondered: what happens when several AIs have to reach an agreement?

    2026 · Three Walls · MSc in Applied AI (UC3M)

    Can agents that don't trust each other negotiate?

    Three factions, Coloso, Enjambre and Errante, separated by walls. All three builders need stone at the same time; if their quartermaster has none, it leans out of the window in the wall and bargains with another faction: bread for stone. BDI agents (beliefs, desires and intentions) that talk to each other with FIPA protocols, built in GAMA.

    What stuck with me: in a system with many agents, the hard part isn't each agent but what happens between them.

    • GAMA · GAML
    • BDI agents
    • FIPA protocols

    With Santiago Herrón Mulet and Keneth Guillermo Ruiz Martínez. The code lives in a private repository of the master's.

    Three factions separated by walls trade bread for stone Coloso Enjambre Errante B Q Q B Q B

    B builder · Q quartermaster

      one sample round; the real model runs in GAMA

      …and I don't know the next question yet. Shall we look for it together?

      drag the magnifier ↓

      Where I've been

      November 2026

      Right now

      I work

      at Cabify, on the Pricing team: services in Elixir and Go, with Docker, Kubernetes and Grafana.

      I study

      the MSc in Applied Artificial Intelligence at UC3M.

      I'm learning

      reinforcement learning, by teaching a bot to play Rocket League.

      the last question

      What's yours?

      If you have a project, a question or just feel like talking about AI, write to me.

      diegopikeras [at] gmail [dot] com