Marcel Tomàs Bernal

• First-author paper at AISTATS 2026, "Breaking Data Symmetry is Needed For Generalization in Feature Learning Kernels" (preprint available upon request) • Designed end-to-end experiment pipelines using PyTorch with dataset versioning, automated evaluation, monitoring, and reproducibility using W&B. • Built distributed training workflows on an 8× NVIDIA H200 multi-GPU cluster, supporting large-scale empirical sweeps of 800+ models per study. • LLM-style transformer experiments, comparing GPT-2 and a simplified GPT, analyzing attention versus depth. • Led empirical and theoretical ML research on neural networks and feature learning, focusing on training dynamics, feature learning and generalization. • Collaborated with PhD researchers, bridging theoretical modeling with production-level ML experimentation.

• Conducted a data analysis project during my 3rd year of university in collaboration with the Barcelona City Council. • Built and maintained a multi-source data pipeline integrating 12+ heterogeneous public and private datasets, including ingestion, cleaning, normalization, and temporal-spatial alignment using Python (Pandas). • Designed a time-series prediction framework for citizen complaints, engineering adaptive thresholds to account for seasonality and population size. • Trained and optimized CatBoost on tabular data, achieving +15.8% F1 and +14.5% balanced accuracy over baseline by refining temporal granularity and introducing custom data augmentation techniques. • Deployed an interactive Streamlit dashboard to visualize insights and support data-driven decision making.