Hi! I'm Juan
This year I finished my master's degree in AI Systems at the University of Trento, Italy. During my time there I worked on a few research projects on multimodal models, neuro-symbolic models and alignment between a language model and the human brain. This experience and all the new published research that we see around language models make me very excited to continue working in research. From the projects, especially, from my thesis, I developed new interests that I want to explore:
- post-training and evaluation of language models
- test-time adaptation
- Coding LLM agents
At the moment I'm working on post-training a Recursive Language Model, introduced by Alex Zhang, and learning about GPU kernels to optimize inference of models.
I would be very happy to discuss any of these topics if you are interested. Feel free to reach out to me via email.
research and theses
-
A sampling-based evaluation framework that separates what an MLLM knows from what it reliably says.
-
A DeepProbLog solver scores 0.91 answer F1 on Raven's Progressive Matrices while two of its three concepts stay at chance.
-
Sparse concept-specific neurons in GPT-2 align with fMRI data better than the dense embeddings around them.
engineering work
-
Agentic workflows for one of Spain's largest livestock cooperatives, cutting document validation from 30 minutes to under 5.
-
13 open-source log parsers compared on accuracy, throughput, memory and tuning effort.
robotics and competitions
I also co-led the university robotics team with Jared and competed in RoboCup. My robotics work includes:
-
Six months on a cognitive robot platform: modular Gazebo description, a drift-correcting physics plugin, and ROS 2 control servers.
-
Generating six-degree-of-freedom grasps for the TIAGo robot, built for the LASR team and RoboCup.
I have also taken part in the NASA International Space Apps Challenge, the Hub Innovazione Trentino Challenge, SciRoc, and RoboCup@Home Education, where our team received a Gold Award.