Four-month machine learning research project analysing regression models on the GAIA DR2 stellar dataset to predict stellar parameters such as brightness, mass, and metallicity.

After performing a literature review, a research gap was identified around ML-based stellar brightness prediction. Multiple regression models were implemented and statistically evaluated using F-tests and Pearson correlation (α = 0.05).

Best models achieved ~99.6% accuracy. Polynomial regression and random forest performed strongest, while SVR showed higher error.

Voice-first space-themed assistant integrating live APIs, NLP, and serverless backend infrastructure.

Built using Python, Flask-Ask, AWS Lambda, DynamoDB, REST APIs, and SSML. Tested locally on a Linux homelab using Docker and Raspberry Pi.

Currently extending into an ML pipeline for predicting movements in space-sector equities using NLP sentiment analysis.

Computer vision-powered tennis platform built at Hack48 that analyses biomechanics and swing footage to generate personalised movement insights.

Combines dominant side detection, flexibility profiling, injury history, and CV-based swing analysis to adapt technique to an individual's body.

Reached 1,000+ users within 24 hours of launch.

Interactive web simulation showing how increasing tree canopy coverage can reduce Sydney's urban heat islands.

Users can draw trees, simulate temperature change, upload community readings, and view live heatmaps.

Built with HTML, CSS, JavaScript, Leaflet.js, and OpenWeather API.

World-first platform integrating real-time NASA satellite data (SWOT, MODIS, PACE) with AI to predict shark foraging behaviour with 87% validated accuracy.

Includes a novel 4D spatiotemporal Shark Foraging Index and a revolutionary non-invasive gastric biosensor for validation.

Aims to improve marine conservation and reduce shark-human conflict globally.

Currently in development.

Built at the Harvard HSIL Hackathon, Mercy is an AI-powered automated triage agent designed to reduce emergency room wait times and lighten the workload on triage nurses. The system reduced both metrics by over 40% in testing.

A multilingual chatbot collects preliminary diagnostic data from patients before they reach a nurse. It uses a custom LLM as its data-collection backbone, asking targeted clinical questions and dynamically updating its decision tree so each follow-up question becomes progressively more specific to the patient's presenting problem.

Built with Python and a custom LLM pipeline. The frontend is served via HTML and the backend handles the conversational logic, decision-tree updates, and triage severity scoring in real time.