AI Research Engineering · Agentic AI · Scientific Machine Learning
I'm a PhD researcher at the University of Surrey, working at the intersection of AI agents, scientific machine learning, and computational research.
I build AI systems for complex technical and scientific problems, from LLM-powered research agents to learned models of physical dynamics. My focus is on combining research with practical engineering, rigorous evaluation, and reproducible results.
-
Prior or Feedback? What an LLM Uses When Adapting Neural Operators: Investigates LLM-guided adaptation of Fourier Neural Operators across PDE regimes. Benchmarks against random search and TPE under matched experimental budgets, with controlled interventions to distinguish scientific priors from experimental feedback. Accepted at the NeurIPS 2026 AI for Science Workshop.
-
Research Brief Agent: An end-to-end LLM research agent built with FastAPI, Qdrant, and Langfuse. Retrieves papers from arXiv, reads full-text evidence, and produces citation-verified technical briefs. Includes streaming, execution tracing, and CI-based scenario evaluations. Achieved zero fabricated citations on the core evaluation suite, at approximately 30 seconds and $0.10 per brief.
-
Differentiable Hybrid Closure for 2D Turbulence (walkthrough): Couples a PyTorch neural closure with a differentiable JAX spectral solver, training directly through simulation rollouts. Reduced 500-step rollout error by 71.7% relative to the standalone solver and 15.7% relative to a matched offline-trained baseline, across 32 test trajectories. Tesseract Hackathon 2026.
-
Neural ODE Surrogates for Black Hole Binary Evolution: Developed parameterised neural ODE surrogates for black hole binary dynamics using N-body galaxy merger simulations, enabling efficient prediction of orbital evolution and merger timescales. First author, accepted to MNRAS.
-
Inverse Parameter Estimation for PDEs: Combines solver adjoints, physics-informed neural networks, and learned posteriors for parameter inference from sparse, noisy measurements. Implemented as interchangeable JAX and PyTorch components. Honourable Mention, Tesseract Hackathon 2025; Tesseract Forum 2026.
ML & Scientific Computing: Python · PyTorch · JAX · Neural Operators · Neural ODEs
Agents & Backend: LLM Agents · FastAPI · Qdrant · Langfuse
Infrastructure & Tools: Docker · GitHub Actions · SLURM · Tesseract