Vitor Sousa

A little context

I studied Information Systems Engineering at the University of Minho. I’m currently a Senior Data Scientist at Wellhub; before that I was at Farfetch, where I worked on recommendation and size-prediction systems. This site is about my independent work and research that has already been published — not about my employers’ internal systems.

What I find interesting

• Evaluation. What a metric actually captures, and where an optimized number and the behaviour we wanted come apart.
• Representation and model internals. What attention, positional encodings, and normalization are doing at the level of shapes and gradients — see Transformer Internals.
• Reinforcement learning for language models. Policy optimization from PPO to GRPO and its variants — see the RL for LLMs series.
• Personalization. Recommendation and contextual bandits, and how to evaluate them offline — see the contextual bandits series.

How I work

I learn by moving between theory, implementation, and experiments. A result isn’t finished for me until I can explain the assumptions, check the implementation, and say what the evidence does—and doesn’t—support.

Books and papers help me ask better questions; the code and tests help me see where my understanding is incomplete. I work through topics in a deliberate order and move on when I can show I understand something — by deriving it, implementing it, and testing it — rather than when I’ve finished reading about it. What I publish is meant to show what I understood and built, with its limits stated.

In practice that means the labels matter: an explainer is a synthesis of known work, an engineering note records a decision or a bug, and a project only says Completed when the evidence supports it. The Notebook holds the smaller, rougher observations along the way.

The references I return to most:

• Tivadar Danka’s mathematics for machine learning, by topic — linear algebra, calculus, probability — when the implementation work needs a concept.
• Understanding Deep Learning (Simon Prince), for deep learning concepts when they matter to the work in front of me.
• Reinforcement Learning: An Introduction (Sutton & Barto), for the RL that post-training runs into.

The longer list is on Resources, and the tools I use are on Uses.

Earlier work and contact

At Farfetch I was part of the team behind Tailor, a size-recommendation approach published at the FashionXRecSys 2023 workshop and co-authored with Alexandre Candeias, Ivo Silva, and José Marcelino. The project page separates the paper’s reported results from my part in it.

Independently, I’m working on Learning with Verifiable Rewards: its model and training foundations are built and tested, and the study itself is paused before its first SFT run. All projects, with their real status, are on Projects.

For a conversation about ML research, evaluation, or something you read here, LinkedIn is the best place to reach me.

GitHubLinkedInCVTailor paper (arXiv)

Last updated 31 October 2026