Problems I've worked on.
Organised by problem domain, not job title. The common thread: hard engineering problems where depth of understanding makes the difference.
Container Terminal Optimisation
A container terminal handles thousands of containers daily. Each one needs to be placed on a vessel, stacked in a yard, and moved by equipment, all while respecting vessel stability constraints, departure deadlines, crane reach limitations, and the physical reality that you can't access a container buried under three others without moving them first.
The container stowage planning problem is NP-hard. Traditional approaches use hand-crafted heuristics that work adequately for small vessels but degrade rapidly as scale increases. I've built reinforcement learning systems that learn stacking and stowage policies from experience, treating the terminal as a sequential decision-making environment.
What makes this work different from typical RL research: every decision has to respect hard physical constraints, the action space is combinatorial, and the system has to operate in real time alongside human planners. The gap between a policy that works in simulation and one that terminal operators actually trust is where most of the engineering lives.
Results
- Multiple RL-based optimisation algorithms deployed and in acceptance testing at major European container terminals
- Systems built to handle hundreds of thousands of real container placement decisions daily
- Published academic paper(s) in collaboration with academic partners
- Custom simulation environment incorporating crane scheduling and realistic operational constraints
- Vessel stowage planning
- Yard stacking optimisation
- Equipment allocation
- Crane scheduling
- Deep reinforcement learning
- Combinatorial optimisation
- Constraint-aware policy learning
- Sim-to-real transfer
AI-Driven Medical Imaging
My PhD at TU Eindhoven focused on making ultrasound imaging faster and more autonomous. The core challenge: ultrasound produces images through a complex chain of signal acquisition, beamforming, and reconstruction. Each step involves trade-offs between speed, resolution, and image quality. I used deep learning to improve these trade-offs.
One line of work applied deep proximal learning, which embeds physics-based signal processing models into neural network architectures, to achieve high-resolution ultrasound images from far fewer transmissions than conventional methods. This isn't just an academic improvement; fewer transmissions means faster frame rates, which matters for real-time cardiac imaging.
Key Contributions
- Deep proximal unfolding for IVUS image recovery from under-sampled data (ICASSP 2022)
- High-resolution plane wave compounding with ~8.2% PSNR improvement using just 3 transmissions
- Co-authored comprehensive survey on ultrasound signal processing with deep learning (25 citations)
- Contributed to COVID-19 lung ultrasound detection (IEEE TMI, 526 citations)
- Plane wave compounding
- Intravascular ultrasound
- Shear wave elastography
- Lung ultrasound (COVID-19)
- Deep proximal learning
- Model-based deep learning
- Deep reinforcement learning
- Compressed sensing
Earlier Research & Experience
Optimisation of stellarator magnetic geometry for increased critical gradients of trapped-electron modes, using proxy relations implemented in STELLOPT.
STELLARATOR OPTIMIZATIONOptimisation of stellarator magnetic geometry for increased critical gradients of trapped-electron modes, combining theoretical physics with multi-variate optimisation.
STELLARATOR OPTIMIZATIONApplied machine learning and big data to decrease simulation time for complex phenomena in fusion plasmas, enabling potential real-time control of plasma evolution in fusion devices.
MACHINE LEARNING FOR FUSION PLASMASResearched chalcogenide glasses, reviewed patent papers, and collaborated on scientific publications.
Fernandes et al. ‘Thermal stability and crystallization kinetics of Bi doped Si₁₅Te₈₅₋ₓBiₓ chalcogenide glassy alloys.’ Materials Today: Proceedings 5.8 (2018): 16237–16245.
CHALCOGENIDE GLASSES RESEARCHBefore research and technology. Trained in decision-making under uncertainty, operating in complex environments, and working within tight constraints. Skills that turn out to be surprisingly relevant to deploying AI in industrial settings. I was also awarded an honorary commission in the French army during this period.