this is what I do:
look inside models like this one
Opening up neural networks to understand what they do
click any word to see how Gemma reads it
pushesinhibitsyour word, ␣do: 19% (rank 2)
hover a node to see its features
steerand {n} moreI'm an AI researcher working where AI alignment, mechanistic interpretability and deep learning meet. For three years at the Applied Artificial Intelligence Lab (LIA2) of the University of Vigo I used causal and correlational graphs to analyse neural activations and map how large language models are aligned inside, and designed my own benchmarks to evaluate how they behave from the outside.
Research is only half of it. I design and deploy full-cycle AI applications and run the infrastructure under them, from on-premise Proxmox and OpenStack to Azure and Google Cloud, with MLOps, DevOps and CI/CD practices so that what works in a notebook also holds up in production.
Before the lab I led the AI team of UVigo Motorsport, turning a Formula Student car into a driverless vehicle that finished in the top 10 at Formula Student UK in its debut. Today I lead GDG Ourense, teaching and organising courses, talks and hackathons. With a master's in Computer Engineering finished, I want to contribute to research teams working on safe, capable and aligned AI systems.
Papers, preprints and datasets I have authored or co-authored.
LIA2 · Universidade de Vigo · 2026
Mechanistic interpretability framework for LLMs. It generates interventional datasets through activation patching, summarises the effect of each patch as a sparse directed graph, and compares those graphs with classical and quantum kernel methods to build a similarity geometry over circuits. Code behind the paper “Patch-Effect Graph Kernels for LLM Interpretability”.
Activation patching on GPT-2 and toy transformers • One sparse graph per slice of prompts • Weisfeiler-Lehman and quantum kernels • Causal validation on a disjoint split
Research tooling · 2026
Streaming Transformer Activation Pipeline: a high-throughput framework for capturing the internal activations of large language models while they run. Hooks feed an async queue, compression and a ring buffer before an async writer, with the goal of keeping at least 90% of baseline inference speed while streaming several GB/s to disk.
Full-tensor, sampled, attention-only and MLP-only hooks • FP16 downcast plus LZ4 or Zstd compression • Per-token checkpoints and resumable captures • Documented stream format for offline analysis
Quantum computing · 2025
Drone route planning as a quantum optimisation problem: a simplified travelling salesman problem is encoded as a Hamiltonian and solved with QAOA on the Qibo framework, with optional GPU execution through its CUDA backend.
TSP encoded as a Hamiltonian • QAOA with configurable depth and optimiser • GPU execution via Qibo's CUDA backend
Personal project · 2025
Local retrieval-augmented generation to fill in documents automatically. Everything runs on the user's machine, so the personal or business data used to complete a form never leaves it.
Fully local RAG pipeline • Customisable document templates • Released under AGPL-3.0
LIA2 · Universidade de Vigo · 2025
Web front end for a digital twin of a smart poultry farm, built alongside the PoultryONT ontology to turn farm data into something a manager can read at a glance.
Dashboard for farm monitoring • Companion to the PoultryONT ontology
Software engineering · 2026
A small e-learning platform with full CRUD for platforms, courses, students, enrolments and lessons: a Spring Boot REST API with business rules enforced on the back end, and a React front end.
REST API with Spring Boot 3 and Spring Data JPA • Reusable lessons shared across courses • React 18 front end with React Router
Personal project · 2025
A small Rust program that reads data from IoT devices and persists it to disk.
The tools I work with, placed where a language model puts them internally: what Gemma sees as similar ends up close together.
Universidade de Vigo · 2024–2026
Master's thesis: Patch-Effect Graph Kernels for LLM Interpretability.
Universidade de Vigo · 2020–2026
Bachelor's thesis: a control and management system for poultry farms, built on an information architecture that enables a digital twin powered by data processing and AI.
Quantum Computing Master Class
Business Analysis & Process Management
LangChain Chat with Your Data
Data Analysis on Google Cloud
Introduction to Large Language Models
Intro to Deep Learning
Machine Learning Onramp
Machine Learning with MATLAB
Azure AI Fundamentals: Introduction to Artificial Intelligence
Building Machine Learning models
Getting and transforming data with Power BI
Preparing data for analysis with Power BI
Data Warehousing Workshop
Introduction to Quantum Computing
Agile Transformation of Organisations
Data Infrastructure Applied to Transport
Practical project management with Microsoft Project