About

Thesis: Profiling the Cost of KV Cache Offloading in LLM Inference. A low-latency C++ inference profiler that attributes energy cost to CPU–GPU interconnect stalls during KV-cache offload, benchmarked on H200 (PCIe) and GH200 (NVLink-C2C) systems.

Coursework: Performance Programming, Accelerated Systems, Machine Learning at Scale, Advanced Message Passing Programming, Numerical Algorithms, Threaded Programming, HPC Architectures.

Dean's List in Fall 2022 and Spring 2021. Coursework in Computer Graphics, Image Processing, Deep Learning, and GPU-Accelerated Computing.

• Represented the University of Edinburgh against 10 selected international teams.
• Optimised llama.cpp batched inference through source modification and performance analysis under a strict 6 kW power budget.
• Built a distributed parallel file system across 8 AMD EPYC client nodes with a GPU-node metadata backend and direct PCIe-attached NVMe targets, delivering 32 GB/s aggregate multi-client throughput.

• Architected a 3D scan coverage analysis workflow processing terabytes of LIDAR data, identifying unseen regions against CAD models.
• Cut an ML pipeline's processing time by 75% through point-cloud handling optimisations and AWS Batch array jobs.
• Built a cross-sectional analysis tool for mooring-chain photogrammetry, improving measurement accuracy from roughly 15% error to sub-millimetre precision.
• Migrated a legacy Bash data-reconciliation process to Python/Prefect, turning a two-day manual task into a five-minute automated one.
• Promoted from Junior (L3) to Level 4 Engineer, later serving as technical team lead for three developers.
• Built proof-of-concept tools in company hackathons, including mesh generation from TLS data, octree compression, and LIDAR scan-quality metrics (Abyssathon 8 winner).

Taught weekly WebGL recitations covering GLSL, the graphics pipeline, Phong shading, mesh loading, and advanced topics including tone mapping, parallax, and shadows.

A matrix multiplication library built around the Factory, Registry, and Meyers' Singleton patterns, with Eigen, tiled CUDA, and cuBLAS backends selected through a JSON config. New backends register themselves without modifying the factory.

A CUDA-accelerated ray casting system for point-cloud intersection testing, using Morton encoding and binary search optimisations.

Compared cuSPARSE CSRColor against the Brook-Vizing randomised colouring algorithm for preconditioning iterative SpMV, reordering elements to launch CUDA kernels in parallel.