Cover Letter

I recently completed a Master's degree in Artificial Intelligence within the Systems Engineering and Computer Science Program at the Federal University of Rio de Janeiro (UFRJ) in May 2025. During my studies, I specialized in Neural Algorithmic Reasoning, a field that investigates how neural network models can be equipped to perform algorithmic computations. My research centered on graph-based models, with a particular focus on how the intrinsic biases of the message-passing scheme in Graph Neural Networks (GNNs) can limit their performance on algorithmic tasks. In my dissertation, I investigated the over-smoothing phenomenon present in message-passing GNNs, especially in heterophilic graph tasks, and analyzed how this characteristic harms the ability of models to preserve node-level distinctions required for correct algorithmic reasoning. To address this limitation, I proposed SpectralMPNN, a hybrid architecture which integrates graph spectral filtering into the message-passing framework, enabling adaptive frequency-based information propagation while retaining the inductive biases that make message passing effective for algorithmic learning.

Over the past years, I have improved my artificial intelligence skills on different academic and professional projects. I have worked as an intern at LampsCo, a spin-off company of the Laboratory of Applied Mathematical Programming and Statistics (LAMPS) at the Pontifical Catholic University of Rio de Janeiro (PUC-Rio), where I worked on the development of a tool for time series prediction and modeling. This experience was extremely valuable as it allowed me to get in touch with traditional statistical models for time series modeling. Following this experience, I joined the IBM Research Lab in Rio de Janeiro, Brazil as a Graduate Research Intern. There, I worked under the Geospatial Modelling group, mainly with fine-tuning tasks for vision foundation models pre-trained on remote sensing data. It has been of fundamental importance to my personal development, as I had the opportunity to work with vision foundation models and get in contact with how they are pre-trained and fine-tuned on an implementation level.

Most recently, following the completion of my Master's, I joined Avra, a Brazilian start-up dedicated to developing tools for a deeper understanding of small and medium-sized enterprises. At Avra, we model the Brazilian corporate ecosystem as heterogeneous graphs and apply graph representation learning techniques to extract meaningful structural and semantic representations. My work focuses on self-supervised learning for heterogeneous graph data, with current challenges centered on designing robust foundational models capable of capturing and encoding the complex, multi-relational structure of the Brazilian economic environment. This experience has further strengthened my interest in scalable graph learning methods and foundation models for structured, real-world data.