I started this site in 2020 to look past the hype and understand what actually worked. Six years later, with the rise of AI assistants, I’m creating things I wouldn’t have imagined attempting then. Follow what I was questioning then and what we can build now.
Do Androids Dream of Running Out of Claude Opus 5.5 Tokens While Animating Electric Sheep?
2026 / NOW
AI assistants have changed how quickly we can turn an idea into a working experiment. Some questions remain difficult. Exploring them has become much easier.
RAG / Public administration
An unofficial chatbot that answers questions about procedures with the Spanish administration, citing official sources. And an audit of the links in the 8,000 records of the SIA catalogue: only half of them work.
RAG / Prototype
Building a RAG got cheaper. Evaluating its accuracy and efficiency remains challenging. Notes from a three-day side project answering questions about Spain’s BOE (official state gazette).
Music / Live application
Different recording. Same song?
Recognizing a particular recording is one thing. Recognizing the same composition through a different singer, tempo, key, or arrangement is another. Cover-song identification has made substantial progress, but radically different interpretations and real-world recordings still challenge current systems.
CoverDetector explores that problem using audio embeddings from CoverHunter and LIVI to search for related recordings. Building the application is only part of the experiment: evaluating its matches, understanding its mistakes, and separating musical identity from superficial similarity are where the harder questions begin.
Try it, listen to the results, and see where it falls short.
Can we find relationships between songs by representing them as vectors?
I started with Chromaprint2Vec, using Chromaprint—an audio fingerprinting tool designed to identify recordings—and MusicBrainz, an open music metadata database.
Harmony2Vec explored learned audio embeddings with CoverHunterMPS, an adaptation of the CoverHunter cover-song identification model that runs on a single computer, alongside chord progression tools.
Mapping these representations lets us inspect which recordings end up close together. Some relationships make musical sense to us. Others are harder to interpret. What has the model learned to consider similar?
2021 / UNCOOL DATA SCIENCE
Causal inference · Updated 17 June 2021
What can algorithms tell us about cause and effect? Experiments with automatic causal discovery, and the assumptions behind the answers.
2020 / THE SITE BEGINS
I started this site to question the promises, work through the methods, and make the assumptions explicit. These articles remain in the context in which I wrote them.
Causal discovery · Updated 22 October 2020
Testing causal discovery methods against examples where the answer seems obvious—and finding that it isn’t so simple.
Statistics · 2 August 2020
Revisiting the fundamentals through health expenditure and life expectancy: model assumptions, panel data, and what the residuals reveal.
COVID-19 / Analysis & simulations
Notes on test errors, bias, and interpreting pandemic data, preserved in their original context.
2019 / BEFORE THIS SITE
Computer vision / Workshop
“Agárrate que vienen curvas.” Workshop materials comparing deep learning and classical machine learning in a self-driving car simulator, based on the Udacity project and NVIDIA’s convolutional model.
2018 / HACKATHONS
Hackathon / TheAlfredoLambdas
A team prototype exploring visual estimates of representation in event and organization imagery.
Hackatrips / Minube & Fitur
A TheAlfredoLambdas team project exploring personalized, gamified travel experiences with augmented reality.
2017 / BERLIN
TechCrunch Disrupt / Berlin
A team hackathon prototype using computer vision to play with the question of getting into Berlin’s exclusive nightclubs.
