Omnilingual team at FAIR. Core contributor to Omnilingual SONAR, where I led the extension of the embedding space to 4,000+ language varieties with a teacher-student distillation framework, reducing cross-lingual similarity error 15x across 1,560 languages. arXiv paper
Core contributor to Omnilingual MT, the first MT system benchmarked on 1,600+ languages, where I built BLASER 3, a massively multilingual MT quality estimation model. Outstanding Paper Award at EMNLP 2026. EMNLP 2026 paper
Continued as a short-term contractor until November 2025 to finalise both projects.
Audio-visual representation learning: a sequential contrastive framework over non-aggregated representations, with up to 3.5x relative improvement in retrieval recall. ICASSP 2025 paper
Contributed to a masked generative video-to-audio model with high audio quality and temporal synchronicity. ECCV 2024 paper
Grade: Excellent, with cum laude distinction.
Thesis: Robust and Data-Efficient End-to-End Speech Translation: Segmentation, Cross-Modal Alignment, and Prosody-Aware Evaluation. → Thesis PDF | UPCommons
Robust segmentation of long-form audio (SHAS), segmentation-based data augmentation (SegAugment), zero-shot speech translation through cross-modal alignment (ZeroSwot, charSONAR), and whether speech translation models use prosody (ContraProST).
Supervised by José A. R. Fonollosa and Marta R. Costa-jussà. Committee: Jan Niehues, Javier Hernando and Elizabeth Salesky.
Grade: 8.3/10, cum laude.
Thesis: Complex Question Answering by Pairwise Passage Ranking and Answer Style Transfer, supervised by Dr. P. Ren. → Thesis PDF
Grade: 7.7/10. Duisenberg Honours Programme in Quantitative Risk Management.
Thesis: An Ensemble of Recurrent Neural Networks for High-Frequency Stock Market Classification, supervised by Dr. S. Borovkova, which led to a publication in the Journal of Forecasting. → Thesis PDF
Multilingual and multimodal representation learning, machine and speech translation, MT evaluation and quality estimation, low-resource NLP
Teacher-student distillation, contrastive learning, cross-modal alignment, sentence embeddings, LLM continual pre-training
Python, PyTorch, fairseq2, HuggingFace, large-scale distributed training (FSDP), Slurm
Coordinated the lectures and the practical labs for the sequence-to-sequence modeling and NLP parts of the course. → Course Page
Supervision of practical labs, assisting students with python and ML questions, grading assignments and final exams. → Course Page
Supervision of practical sessions, grading assignments and final exams. → Course Page
BSc Thesis Supervisor for Pol Rosinés Pozo → Thesis PDF
BSc Thesis Supervisor for Miquel M. de Morentin Cardoner → Thesis PDF
Research Project Supervisor for Pau Lozano García
This project features an AI DJ system designed to perform a collaborative back-to-back set with a human artist. Uniquely, the system is sound-agnostic; it uses a language model to understand the sentiment of YouTube comments to find songs with a similar "feeling" to the recent track history. The project was showcased at the various events including S+T+ARTS Festival in Barcelona and Sonar+D in Lisbon.