Technical Projects

Parallelised analysis of the COMP90024 Twitter dataset on the University of Melbourne SPARTAN HPC cluster. Using the provided created_at and pre-computed sentiment fields, I partitioned the JSON across workers to compute the happiest hour/day (max aggregate sentiment) and most active hour/day (max tweet count). Jobs were submitted via SLURM and scaled across 1×1, 1×8, and 2×8 core/node configurations, recording end-to-end runtime and speedup on the shared system.

Built a fast, accessible portfolio focused on clarity over flair. Problem: keep content accurate and maintainable. Approach: React Router 7 + TypeScript + Tailwind with reusable cards and conditional media; Terraform for AWS learning, deployed on Vercel for simplicity. Outcome: consistent UI, standardized actions, and easy content updates.

Built a fact-checking system for the COMP90042 climate-claims task: given a claim, retrieve relevant evidence passages from a large corpus and predict one of {SUPPORTS, REFUTES, NOT_ENOUGH_INFO, DISPUTED}. The pipeline uses Sentence-BERT + FAISS for dense retrieval, cross-encoder re-ranking, and Gemma-3-1B-it for claim classification, delivering a strong recall lift over the provided baseline on a 1M+ document set within Colab-scale constraints.

Built a real-time pipeline on the Melbourne Research Cloud: Fission harvesters on Kubernetes pulled VicRoads traffic, BoM weather, and EPA air-quality feeds, transformed them, and loaded Elasticsearch indices. I analyzed travel time and pollutant trends (PM10/PM2.5, etc.) across freeway segments in Jupyter, highlighting relationships between congestion and air quality and proposing improvements to reduce cold starts and streamline data flow.

A Chrome extension for reading Korean web novels where text is embedded in images. I built a selection mode to pick images on the page, run OCR with Tesseract.js (bundled locally to satisfy Manifest V3), then batch-translate the extracted Korean into natural English using the OpenAI API. Translations open in a dedicated results window with saved history and adjustable font/size, and the API key is managed via Chrome local storage.