projects & portfolio.

Built a machine learning system to predict antimicrobial resistance patterns from field-collected surveillance data. Covers the full pipeline: data collection, preprocessing, and supervised ML modelling to support clinical decision-making in resource-limited settings.

The first formal survey of safety risks in multi-agent LLM pipelines — covering constraint drift, prompt injection between agents, identity collapse, monoculture vulnerabilities, and emergent deception. Fills a gap that Anthropic, OpenAI, and DeepMind have not formally addressed yet.

Built an AMR drug-resistance prediction system from the ground up — field data collection, preprocessing pipeline, and supervised ML modelling. ReAct is a global network operating across 90+ countries focused on antibiotic resistance. First real-world ML deployment experience, working alongside public health specialists.

A fully local AI assistant built on open-source LLMs — no cloud APIs, no data leaving the machine. Handles task orchestration, retrieval-augmented generation, and agentic workflows. The foundational project that led directly to my research on multi-agent safety.

Mech interp has cracked open single models — this paper asks what happens to those circuits when agents delegate, pass memory, and use tools. Surveys the current state of the field and proposes a research agenda for extending interpretability to multi-agent settings.

Selected as a Junior Academy member at the New York Academy of Sciences. Currently developing MineQuanta — a project exploring quantum computing approaches to optimise bio-mining processes. One of a small global cohort of students working on interdisciplinary STEM challenges under the NYAS framework.

A full-stack automation tool that scrapes Australian job boards and applies to casual roles automatically — Easy Apply jobs are auto-submitted, complex ones are pre-filled for manual review. AI-generated cover notes via Claude API.

AMR will kill 39 million people by 2050. This systematic review maps how ML models use genomic sequencing, clinical records, and surveillance data to predict resistance — then diagnoses why deployment stalls. Covers data fragmentation, geographic bias, interpretability gaps, and the clinician trust problem.

A progressive web app for tracking academic deadlines, grades, and unit progress across a university semester. Built out of personal need during Semester 1 2026 — installable, offline-capable, and deployed on Vercel.