Vitor Sousa · Research & engineering
Learning systems,
from the inside out.
I’m Vitor, a Senior Data Scientist at Wellhub, previously at Farfetch. I work at the intersection of machine learning, software engineering, and careful measurement.
At Farfetch, I helped build size recommendations at marketplace scale. Independently, I build models from raw tensors and write about transformer internals, post-training, and evaluation — with code and tests alongside the ideas.
Research depth.Production discipline.
- 4M+ customers · 190 countriesFarfetch marketplace context
- FashionXRecsys 2023Co-authored Tailor paper
- A Transformer · 112 testsIndependent, from-scratch implementation
Code, methods & evidence
Selected work
Learning with Verifiable Rewards
A research-engineering study of whether GRPO earns its cost over a properly built SFT baseline on GSM8K, and where it fails. Model and training foundations are built and tested; the SFT baseline is in progress; GRPO, ablations, and findings are not yet run.
Artifact · PyTorch implementation + test suite
Tailor at Farfetch
Size recommendation as sequence classification. Public Farfetch work using LSTMs and attention to model evolving fit preferences.
Paper · FashionXRecsys 2023
Notes from the work
Selected writing
A Transformer from Raw Tensors: What 112 Tests Taught Me
The Phase 1 close-out of rlvr-from-scratch: 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.
After GDPO: A Map of the GRPO Family
GRPO didn't fragment into a zoo. It exposed a handful of design axes — token vs. sequence granularity, clip width, length and std normalization, KL, advantage normalization, sampling — and every 'new' algorithm is a different point in that space. A practitioner's map of Dr. GRPO, DAPO, GSPO, GDPO, and the frontier, with a decision guide for when to reach for each.
New here? Find a reading path, or go deeper with Transformer Internals.