I build production ML systems where learning, decision-making, and evaluation matter.
I’m Vitor Sousa, a Senior Data Scientist at Wellhub on the GenAI & Engagement team, where I build production machine learning systems focused on personalization, decision-making, and user engagement. Previously at Farfetch, I built recommendation and size-prediction systems serving 4M+ customers across 190 countries — deep learning from scratch, learning-to-rank, and a published paper at ACM RecSys. This site goes beyond the day job — it’s where I investigate research questions, build systems from first principles, run experiments, and write about what the evidence actually shows — lately: post-training, evaluation, and how learning systems behave, fail, and improve.
13 articles · 5 projects
Featured work
Learning with Verifiable Rewards
An empirical study of GRPO, controlled ablations, uncertainty, and failure modes in reinforcement learning with verifiable rewards.
A Transformer from Raw Tensors: What 112 Tests Taught Me
The model-foundations close-out of Learning with Verifiable Rewards: a decoder-only transformer built from raw tensors — attention with RoPE, RMSNorm, SwiGLU, an incremental KV-cache, four sampling modes — and the three properties that are easiest to get silently wrong. A case study in verification, not assembly.
Beyond the Vibe Check: A Systematic Approach to LLM Evaluation
Stop relying on gut feelings to evaluate LLM outputs. Learn systematic approaches to build trustworthy evaluation pipelines with measurable metrics, proven methods, and production-ready practices. A practical guide covering faithfulness vs helpfulness, LLM-as-judge techniques, bias mitigation, and continuous monitoring.
Selected writing
See also: Foundations →A Transformer from Raw Tensors: What 112 Tests Taught Me
The model-foundations close-out of Learning with Verifiable Rewards: a decoder-only transformer built from raw tensors — attention with RoPE, RMSNorm, SwiGLU, an incremental KV-cache, four sampling modes — and the three properties that are easiest to get silently wrong. A case study in verification, not assembly.
Beyond the Vibe Check: A Systematic Approach to LLM Evaluation
Stop relying on gut feelings to evaluate LLM outputs. Learn systematic approaches to build trustworthy evaluation pipelines with measurable metrics, proven methods, and production-ready practices. A practical guide covering faithfulness vs helpfulness, LLM-as-judge techniques, bias mitigation, and continuous monitoring.
Selected projects
Learning with Verifiable Rewards
An empirical study of GRPO, controlled ablations, uncertainty, and failure modes in reinforcement learning with verifiable rewards.
rlvr · grpo · evaluation · reinforcement-learning · experimentation · pytorch
Tailor: Size Recommendations at Farfetch Scale
Sequence classification models for personalized size prediction in luxury fashion — LSTMs, attention mechanisms, and a published paper at ACM RecSys 2023.
recommendation-systems · deep-learning · pytorch · sequence-models · attention · production-ml · ab-testing
RAG System with LlamaIndex, Elasticsearch & Llama3
Local-first RAG pipeline with hybrid search: BM25 + dense retrieval on Elasticsearch, LlamaIndex orchestration, and Llama3 for generation. Evaluated with RAGAS metrics across chunking strategies and retrieval configurations.
Elasticsearch · LlamaIndex · Llama3 · RAG · Vector Search