I build and research machine learning systems — from deep learning and quantum machine learning through to the data pipelines that feed them. Master of Data Science (Professional), Deakin University.
I'm a data scientist and machine learning engineer with a foundation that spans modelling and research — deep learning, reinforcement learning, generative models — and the applied data engineering that turns experiments into working systems. My academic and research work has centred on deep learning and quantum machine learning, while my industry experience has put me deep into data pipelines, RAG architectures, and rapid prototyping across several teams.
I'm completing a Master of Data Science (Professional) at Deakin University, specialising in Cyber Security and Deep Learning, and I'm drawn to research-driven teams, consultancies, and early-stage products where greenfield modelling and fast iteration are valued. I care about getting models out of the notebook and into something people can actually use.
Designing and training neural networks for vision, sequence, and generative tasks — and reasoning about why they behave the way they do.
Taking models from prototype toward production — building LLM & RAG applications and the scaffolding around them.
Building the pipelines and warehouses that make modelling possible — ETL, integration, and clean, queryable data.
Academic research at Deakin exploring where quantum approaches meet learning algorithms — an emerging field I follow closely.
A controlled study adapting DistilGPT2 to the formal register of two policy reports. I benchmarked baseline perplexity, then compared full fine-tuning against a low-rank LoRA adapter — the adapter recovered most of the domain shift while training just 0.18% of the model's weights (147K vs 82M parameters), trading a little accuracy for a large drop in cost.
A YOLOv8 object detector for fabric-stain quality control, motivated by a railway-linen inspection use case. With in-domain data scarce, I bootstrapped the detector on a larger public fabric-stain corpus and built the full annotation-to-training pipeline myself — converting Supervisely JSON labels to YOLO format and handling the train/val split. Reached mAP50 ≈ 0.85 on a 96-image validation set.
A hybrid quantum-classical reinforcement learning project that injects a parameterised quantum circuit (angle-encoded, ring-entangled) directly into Stable-Baselines3's A2C as a custom feature extractor — enabling a clean, single-codebase comparison of quantum versus classical agents on a satellite channel-allocation task. Built in PyTorch and PennyLane with backprop differentiation for fast CPU simulation, and circuit depth exposed as a single sweepable parameter for studying the expressivity-vs-trainability trade-off.
Built data-flow and ingestion pipelines integrating voice and gesture recognition for real-time AI-robotics proof-of-concepts.
Developed end-to-end data-integration pipelines for AI-driven applications, including a home assistant and a healthcare companion app.
Architected end-to-end RAG pipelines with LlamaIndex for enterprise AI tools, streamlining unstructured-document workflows.
Designed extraction/transformation pipelines for large-scale e-commerce image and text data, feeding aspect-based sentiment reporting in Tableau.
Built ETL pipelines (Tableau Prep, Pandas) and a relational data model to support e-commerce sales-forecasting analytics.
Built automated data-collection and classification pipelines in Python, integrated with cloud storage.
Validated large-scale data-pipeline integrity (Teradata, SQL) for a major data-lake rehydration project.