AI research engineer
I build LLM agents and train models for physical simulation.
PhD researcher in astrophysics and machine learning at the University of Surrey.
A study of how an LLM uses task descriptions and validation feedback to choose neural-operator fine-tuning settings. I compared search methods under matched budgets and built the accompanying package for recording and replaying runs.
Accepted to the NeurIPS 2026 AI4Science Workshop. Equal contribution with Javier Mora Jimenez.
I built a FastAPI agent that searches arXiv and a Qdrant paper index, reads sources and writes cited briefs. It checks citations, limits tool calls and records traces in Langfuse. Offline scenario tests run in CI.
I trained a PyTorch turbulence closure through a JAX spectral solver. In-process training reduced 500-step rollout vorticity MSE by 71.7% versus the solver alone across 32 test trajectories. A separate demonstration passes gradients through two served Tesseract components.
Solver adjoints and JAX/PyTorch PINNs for Burgers viscosity estimation.
I trained parameter-conditioned neural ODEs on N-body simulations of black hole binaries and tested them on held-out trajectories and an unseen higher-resolution case. Their predictions feed a semi-analytical model of binary merger times.
First-author paper published in Monthly Notices of the Royal Astronomical Society, 2026.
My research models black hole binary evolution with machine learning. I also develop GPU simulation software and differentiable JAX/PyTorch pipelines.
I hold an MSc in Computer Science from Swansea University and an MSci in Physics from the University of Birmingham.
Scientific ML CV