Selected work

Projects

Strongest first. This list will grow as more work is written up.

• 01

  GPU and parallel performance engineering

  COMPX553, 99.05%

  A nine-assignment progression from sequential Java tuning through OpenCL and CUDA to Hadoop and Spark, built on one rule - measure, optimise, and attribute every gain to a specific cause.

  CUDA C++, OpenCL, Java (ForkJoinPool, CompletableFuture), Hadoop MapReduce, Spark

• 02

  EEG flow-state detection - funded summer research

  2025-26

  A 10-week funded research project on decoding flow state from wearable EEG, now continuing into an honours dissertation.

  Python, MNE/DSP, EEG-BIDS, scikit-learn, Optuna, MLflow, LSL/XDF

• 03

  Deep learning under real-world constraints

  COMPX525, 99.14%

  Two PyTorch projects on messy, real-world image and audio data rather than clean benchmarks - one strong result, one well-evidenced negative result.

  PyTorch, transfer learning (DINOv3, BirdNET), LoRA, Optuna, GPU-side augmentation

• 04

  RBFN benchmark and Bayesian inference

  DATAX397, 94.63%

  A custom Radial Basis Function Network benchmarked rigorously against tuned baselines, alongside a from-scratch Bayesian inference study.

  Python, NumPy, custom scikit-learn-style estimators, Optuna, Bayesian inference

• 05

  Streaming ML

  COMPX523, 92.74%

  A memory-budgeted streaming ensemble with drift detection, plus fixes to a streaming KNN implementation's memory growth and interop overhead.

  CapyMOA/MOA, NumPy, concept-drift handling, online ensembles

• 06

  Spectral graph theory report

  COMPX546, 99.28%

  A self-chosen technical report building spectral graph theory from the adjacency matrix to the Fiedler vector, with two original proofs and code-generated figures.

• 07

  whisper-transcriber - local ASR and agent tooling

  A general-purpose long-form-audio transcription pipeline, plus an honest capability finding from a shelved local-LLM cleanup experiment.