Filippo Giovagnini

Hello! I am a PhD candidate in Mathematics at Imperial College London, supervised by Professor Dan Crisan. My research concerns stochastic analysis and numerical methods for fluid-dynamics PDEs and SPDEs, with a particular focus on vortex methods and interacting particle systems.

Working with particle systems, McKean–Vlasov equations and nonlinear PDEs led me to explore modern machine-learning architectures from a more mathematical perspective. I am increasingly interested in understanding the internal behaviour and representations of large models.

Alongside my PhD, I work on applied machine learning projects at Stripe Partners.

I enjoy working where theory informs engineering, and where experiments help refine theoretical understanding.

Selected machine learning work

Kaggle Silver Medal - NeuroGolf 2026

Ranked 40th (top 1.4%) in an AI-agent and ONNX optimization challenge. Built an iterative agent workflow for exact verification, parameter reduction and memory-constrained model design.

Code Competition

Decoder-only Transformer and distributed training

Developed a hands-on PyTorch project exploring causal attention, RoPE, Mixture-of-Experts and distributed data-parallel training. Prepared a local FineWeb byte-token dataset and trained a 150M-parameter model for about 1B byte tokens across four RTX 3090 GPUs, with validation, checkpointing and W&B metrics.

First large-scale run completed · September 2026

Code Interactive results