Selected work

Projects.

Where the mathematics meets the machine — research projects spanning multi-objective optimization, video understanding, medical ML, distribution shift, and benchmarking.

Choquet Adaptive Scalarisation (CAS)

Aug — Sep 2026

SIT316 Linear & Nonlinear Optimisation · Deakin University · Supervised by Prof. Gleb Beliakov

Designed a novel nonlinear multi-objective optimiser that scalarises objectives with the Choquet integral and adaptively updates its parameters from the non-dominated solutions found along the way.
Leveraged the nonlinearity of the Choquet integral to capture non-convex, curved, disconnected, and unevenly distributed Pareto fronts.
Benchmarked on ZDT1–3 and DTLZ1–2 against a linear-scalarisation baseline (NNLS updates): slightly higher hypervolume and more non-dominated solutions, at a higher computational cost.

PythonNumPySciPypymoopyfmtoolsMulti-objective Optimization

Report (PDF) ↗ GitHub repo ↗

Sign Language Recognition with X3D

Aug — Sep 2026

SIT332 Robotics, Computer Vision & Speech Processing · Deakin University

Summarised X3D, an efficient video architecture that expands a small 2D network one axis at a time, and its implementation in PyTorchVideo.
Fine-tuned three Kinetics-400-pretrained X3D variants on AUTSL (226 Turkish signs) at six freezing depths, run as Slurm job arrays on A100 GPUs.
Best accuracy–cost trade-off: X3D-M with two frozen blocks reached 82.3% top-1 at 4.9 GFLOPs per clip, matching full fine-tuning; also tested on a hand-crafted sign dataset.

PythonPyTorchPyTorchVideoVideo UnderstandingSlurm / HPC

Report (PDF) ↗ GitHub repo ↗

Improved DASMcC — Heart Disease Prediction from 12-Lead ECG

Aug — Sep 2026

SIT307 Machine Learning · Deakin University

Attempted to reproduce DASMcC, a published multi-class cardiovascular-disease classifier on PTB-XL, and found its reported 88–93% accuracy was not reproducible.
Traced the gap to SMOTE applied before the train/test split; an ablation leaking data the same way recovered 80.5–88.4% accuracy.
Proposed a 1D-CNN + autoencoder feature pipeline that raised the best macro-F1 from 40.5 to 58.4, and showed tree-based models gain further on uncompressed features.

PythonPyTorchScikit-learnXGBoostCatBoostReproducibility

Report (PDF) ↗ GitHub repo ↗

Out-of-Distribution Data Fitting

Nov 2025 — Feb 2026

ADR Summer Project 2025 · Deakin University

Implemented a range of data-splitting techniques designed to mimic real-world distributionally shifted data.
Benchmarked state-of-the-art ML/DL models under those shifted regimes to probe how generalization degrades off-distribution.
Summarised findings and reviewed the related literature in an academic, reproducible manner.

PyTorchScikit-learnBenchmarkingDistribution ShiftLiterature Review

GitHub repo ↗

OODToolkit — Tabular Regression Benchmarking Pipeline

2025 — 2026

Companion toolkit · Out-of-Distribution Data Fitting

Built an end-to-end, config-driven pipeline that splits tabular datasets, trains regressors, and evaluates them under out-of-distribution train/test regimes.
Engineered geometric and distributional splitters — hyperball, slab, K-means, and covariate-shift — to synthesise controlled distribution shift from any dataset.
Unified eleven classical and neural regressors (Huber, KNN, SVM, tree ensembles, XGBoost, LightGBM, ResNet) behind one interface, with Slurm batch support for HPC runs.

PythonPyTorchScikit-learnXGBoostLightGBMSlurm / HPC

GitHub repo ↗

Linear Programming for Data Science

Jun 2024

PiMA Research Summer Camp 2024

Transformed regression problems into optimization frameworks and clean matrix formulations.
Built linear-regression models with both L1 (linear programming via SciPy linprog) and L2 (closed-form least squares) approaches.
Implemented and evaluated both in Python, comparing error-minimisation strategies on robustness, accuracy, and sensitivity to outliers.

PythonSciPyLinear ProgrammingLeast SquaresRegression

GitHub repo ↗

The lab is always running.

New work in graph neural networks and brain-inspired models is in progress. Check back, or follow along on GitHub.

Follow on GitHub →