Geometric deep learning
Graph neural networks, node embeddings, and learned representations over meshes and point clouds, the structures that physical simulation actually lives on.
GNNsnode embeddingsmesh learningpoint clouds
01
About
I work at the intersection of deep learning and computational fluid dynamics. At Mercedes-AMG Petronas I build aerodynamic ML, from graph-based learning on meshes to 3D transformer surrogates that predict pressure and flow fields directly from geometry.
This autumn I begin a DPhil in Oxford's Visual Geometry Group, where my interests sit where geometry, representation learning, and physical simulation meet.
02
Research interests
3D transformers for CFD
Attention over aerodynamic geometry: surrogate models that approximate the solver, predicting pressure coefficients and flow fields orders of magnitude faster than full simulation.
3D transformersCFDaerodynamicssurrogate models
Learning from imperfect data
Noisy labels, severe class imbalance, and the realities of scientific and security datasets, training models that hold up when the supervision doesn't.
noisy labelsimbalancesecurity MLmalware
Simulation & virtual worlds
Procedural generation, physics-grounded rendering, and visual effects, from real-time fluid solvers to black-hole simulations and small physical test rigs.
Navier–Stokesprocedural genVFXphysical sim
03
Selected publications
04
Path
2026 →
DPhil candidate, Visual Geometry Group
University of Oxford · Prof. Andrea Vedaldi · DLA funded
Beginning doctoral research in 3D and geometric deep learning, bridging learned representations and physical simulation.
2024 to present
ML Research Engineer
Mercedes-AMG Petronas Formula One Team · Aerodynamics / Data Science
Research engineering for aerodynamic ML, graph-based learning evolving into 3D transformer surrogates, trained on NVIDIA H200 clusters. CFD data pipelines, mesh processing, and inference tooling.
2023 to 2024
Master's, Computing (AI & Machine Learning)
Imperial College London · Distinction
Deep learning from noisy and imbalanced security data; research later accepted to ACM CCS.
05
Projects
P-01↗
CFD post-processing pipeline
End-to-end VTK/VTP/VTM processing for wind-tunnel runs: pressure-coefficient mapping across cut-ups and cutdowns, curvature-normal promotion, and interactive scrollable plots of full runs.
PyVistaVTKFastAPIPython 3.11
P-02↗
Regulation intelligence
Local LLM system for parsing and reasoning over FIA Formula 1 technical regulations, with PDF extraction, vector retrieval, and a vision-capable chat interface, all running on-prem.
GemmavLLMWeaviatePyMuPDFDocker
P-03↗
Neural tap anomaly detection
Real-time anomaly detection over sensor tap data, served as a lightweight PyTorch inference service behind a FastAPI endpoint.
PyTorchFastAPICUDA
P-04↗
Stable-fluids lab
A GPU Navier–Stokes solver: advection, vorticity confinement, and a Jacobi pressure projection running entirely in WebGL2. It is the ink behind this page, so move your cursor to disturb it.
WebGL2GLSLfluid sim