Marcelo Correia
Selected work

Projects

Automated Customer ReviewsTeam

NLP Pipeline · 2026

A pipeline that turns 129,000 Amazon reviews into product insights: sentiment classification with fine-tuned DistilBERT and RoBERTa, product clustering with sentence embeddings + KMeans, and BART-generated comparative summaries, shipped as a Gradio app with separate consumer and seller views.

  • 82.97% Macro F1 vs. 49% for a zero-shot baseline, which scored 0% recall on the neutral class.
  • Diagnosed and fixed a silent embedding-truncation bug via chunked mean pooling.
0%
Macro F1 (vs. 49% zero-shot)
0k
reviews processed
  • Python
  • Hugging Face Transformers
  • DistilBERT
  • RoBERTa
  • BART
  • Sentence embeddings
  • KMeans
  • Gradio

Fine-Grained Image ClassificationTeam

Oxford-IIIT Pets · 2026

As technical lead: fine-grained classification of 37 visually similar pet breeds. Benchmarked a from-scratch CNN against transfer learning with ResNet18, ConvNeXt-Tiny and ViT-B/16.

  • 94% test accuracy with the best model (ConvNeXt-Tiny).
  • Debugged shared-transform leakage between train and validation sets.
0%
test accuracy (ConvNeXt-Tiny)
0
breeds
  • Python
  • PyTorch
  • CNN
  • ResNet18
  • ConvNeXt-Tiny
  • ViT-B/16
  • Transfer learning

More projects coming soon…