My name is Giles!
I'm a doctor of physics, and a research-engineer specialising in AI/ML/DL and automated design. I have ten years experience in deep learning and differentiable programming, and its application to cutting edge experiments and problems.
The majority of this experience has been gained in the field of high-energy physics, in which I was an active member of the European Centre for Nuclear Research (CERN) for ten years. During this time I worked on a variety of projects centred around the application of neural networks to statistical data-analysis. This included studying the properties of the Higgs boson, developing original, high-performance algorithms for enhancing signal-to-noise ratios, developing domain-specific deep-learning packages (LUMIN, PyTorch_Inferno), and producing recommendation reports for future experiment constructions.
Naturally, this work required a deep understanding of the underlying physics, strong skills in statistics and data-science, distributed HPC, and staying up to date with the latest developments and techniques in machine learning (ML). My PhD work contributed to publishing the most stringent limits on Higgs boson pair-production using three years worth of data collected at the Large Hadron Collider (LHC) (Nature summary).
Following my PhD, my research shifted to expanding the domain of automated design; in particular, the application of differentiable programming to the optimisation of particle-physics detectors and experiments. Through the use of differentiable physics simulators and analysis/inference chains, an end-to-end pipeline can be constructed, in which parameters of the design (both hardware and software) can be optimised to maximise performance, or the scientific output of experiments, with full analytic considerations of measurement uncertainties.
This technique was successfully demonstrated in the context of muon-tomography, in which the designs of muon-tracking detectors were optimised to maximise the resolution of the reconstructed image in an industrial context. This lead to a variety of publications (main paper, NeurIPS paper, community whitepaper), and an open-source package implementing the full simulation, inference, and optimisation pipeline (TomOpt), for which I was the lead developer. In parallel to these efforts, I was a founding member of the MODE Collaboration, which aims to bring together top researchers from all over the world to develop and apply these techniques to a wide range of problems.
Following three years of post-doctoral research, I moved into industry, working for Tokyo-based deep-tech startups specialising in the application of AI, physics, mathematics, and geometry to automated industrial design and Computer vision, and LLM-based agentic AI for Digital Transformation (DX) and Natural Language Porcessing. My work builds on my previous experience, and has allowed me to continue to grow and develop skills in geometric deep learning, surrogate modelling, constrained optimisation, team collaboration, industrial design, industry-style R&D, industry-level software development, and production-level engineering and deployment.
In my free time I have been continuing some academic research with my old collaborators, and develop personal projects to continue to push my R&D and engineering skills, e.g. agentic AI for MTG deck building