I build foundation language models and multilingual systems. My work spans Llama pretraining and inference, and efficient machine translation.
Work on Llama 3 and 4 at Meta Superintelligence Labs. Focus on pretraining efficiency through distillation, pruning and scaling, plus reasoning and inference optimization. Also clean and deduplicate high-quality multilingual web data.
Built high-performance machine translation systems tailored to each client, designed to run locally for privacy. Developed complete model pipelines covering data bitexting, augmentation, knowledge distillation, quantization and evaluation.
Scaled machine translation pipelines to process 12 petabytes of web data for the HPLT project. Published open-source bitexts and efficient translation models. Worked on CPU and GPU inference optimizations and achieved competitive results in the WMT Efficiency Tasks.
Improved simulations of QCD phenomena for the LHC by updating modeling code for the Herwig and HEJ Monte Carlo frameworks. Implemented key computations for vector-boson scattering in the high-energy limit.
Analytical research and software development for computational simulations of QCD phenomenology at the Large Hadron Collider. Investigated new accuracy improvements (KrkNLO) and analyzed parton shower uncertainties.
First Class Honours.
C++ · Python · Bash · Git · SQL · $\LaTeX$
PyTorch · PySpark · HPC · Linux · macOS
HTML · CSS · JavaScript
English (native) · Italian (basic)