projects
Write-ups of things I've built and worked on — research projects, thesis work, and internships.
-
What Multimodal Models Know but Don't Say
My MSc thesis started as an attempt to make vision-language models more accurate at open-world recognition. It ended up somewhere more interesting: showing that the instruments we measure these models with are not neutral, and that a lenient judge and a strict audit can flip the ranking of the same models on the same data.
By Juan Camacho Mohedano, MSc in Artificial Intelligence Systems, University of Trento. Thesis supervised by Elisa Ricci and Marco Garosi. All figures below are from my thesis and defence presentation unless noted otherwise.
-
Discovering Reasoning Shortcuts in RAVEN
Can a model get the right answer for the wrong reasons? For my Advanced ML project at Trento I extended the rsbench benchmark suite to Raven’s Progressive Matrices, built a neuro-symbolic solver with DeepProbLog, and caught it red-handed: up to 0.91 answer F1 while two of its three concepts were pure noise.
This project builds on the official rsbench benchmark suite for studying concept quality and reasoning shortcuts. I worked on it under the supervision of Samuele Bortolotti, and contributed the RAVEN implementation in PR #3.
-
Do a Language Model's Expert Neurons Think Like a Human Brain?
Language models align with human brain activity during language processing, but why? Is it because every neuron contributes a little, or because a sparse set of “expert” neurons carries the semantics? For my Artificial and Biological Neural Systems project at Trento, I took the expert neurons out of GPT-2 and compared their representational geometry directly against fMRI recordings from nine people.
-
Putting AI Agents to Work in a Livestock Cooperative
Research problems come with a benchmark. Business problems come with a person who spends thirty minutes a day checking documents by hand. During my internship at COVAP I worked on the second kind: building agentic AI workflows to automate document-heavy processes at one of Spain’s largest livestock cooperatives.
-
Which Log Parser Should You Actually Use?
Early in 2025 I took a freelance contract with a deceptively simple brief: a system was producing more logs than anyone could read, and someone had to decide how to make sense of them automatically. The interesting part is that most of the decision sits upstream of where people usually look. Before you can flag a single anomaly you have to turn raw log lines into clean templates, and the parser you pick quietly sets the ceiling for everything downstream. So I did the honest thing and benchmarked the options properly, going from a survey of the methods to a recommendation backed by numbers.
-
Building Software for the MiPA Robot at NEURA Robotics
From January to June 2023, I worked as a Robotics Developer in the R&D team at NEURA Robotics. I spent six months working with MiPA, a robot platform designed for everyday interaction, and moved between simulation, grasping, perception, and the software needed to operate the robot safely.
-
Teaching Robots to Grab Stuff: My BSc Thesis Adventure
For my BSc thesis in Computer Science with AI at the University of Leeds, I worked on 6DoF robot grasping for the TIAGo robot. The aim was to let TIAGo find a sensible way to pick objects up from a table in a simulated room, as part of the kind of service-robot task used in RoboCup@Home.
Thesis supervised by Mehmet Dogar at the University of Leeds and Matteo Leonetti at King’s College London.