WhoamI?
I have too many interests at any given time. They have evolved, devolved and convolved, But what has stayed constant is that I like dabbling in whatever looks interesting and digging deeper into some of them along the way.
I have worked as a data scientist since completing my Master of Data Science at the University of Sydney. I discovered the field after stumbling upon the R Programming course on Coursera while studying for my Bachelor's in Finance, and I later decided to pursue it as a career after getting hooked on Chris Manning's lecture "Building Neural Network Models That Can Reason", the same topic I was researching for my Master's thesis.
My pull towards AI safety came when I bumped into Rob Miles's videos on mesa-optimisers and went down the rabbit hole of finding everything I could, as I had been bored of building variations of the same ML models for some time by then. And the reason I continued on this track was the realisation that the world does not need another person to work on developing and deploying (which was really what I was doing) more and more capable AI models, so I have made it kind of a mission to work on anything that is adjacent to AI risk and safety ever since.
I am based in the sunny Eora Nation (more commonly known as Sydney), Australia, and occasionally work from Sydney AI Safety Space (SASS).
I have been generally focusing on socio-technical AI safety, which means I find myself doing a blend of both technical and policy research at any given time.
Machine Intelligence & Normative Theory (MINT) Lab - Johns Hopkins University
2026-present
Centre for AI Risk Management and Alignment (CARMA)
2026-present
Future Impact Group
2025-2026
Mentorship for Alignment Research Students (MARS) - Cambridge AI Safety Hub
2025-2026
Supervisory Program for Alignment Research (SPAR)
2025
Lumi Finance
2024-2026
Commonwealth Bank of Australia
2022-2023
Allianz Australia
2021-2022
Deloitte Australia
2020-2021
2026 Incoherent Values? Probing LLM Preferences Through Parametric Variation. Tests whether a model’s stated preference still holds when the property the model values is made stronger or weaker and proposes a statistical framework for evaluating value coherence.
2026 Principles and Guidelines for Randomized Controlled Trials in AI Evaluation. Proposes five principles and thirty-three guidelines for running AI human-uplift randomised controlled trials, drawing on practice from software engineering, economics, health sciences and psychology so that study results are transparent, repeatable and comparable.
2026 Mapping U.S. Federal AI Governance Against Sector Vulnerability. Compares the governance coverage of 684 U.S. federal AI governance documents across 14 sectors and 24 AI risk subdomains with how vulnerable experts rate those sectors, and finds where coverage is lacking to inform governments and industries on AI risk related decisions.
2025 An Empirical Study of Power-Seeking Behaviour in a Multi-Agent Setting. Simulates four Gemini 1.5 Flash agents running a shared art gallery for six months to test for power-seeking behavior to compare a collaborative baseline against a scenario where the agents are told to prioritise self-interest.
2026 How to Get Into a Career in AI Safety Research and What Do They Do? A conversation about my journey into AI safety research and my perspective on what it takes to build that career from Australia.
2026 Incoherent Values? Probing LLM Preferences Through Parametric Variation. A talk on whether a model’s stated preference still holds when the feature it values is made stronger or weaker, and on the framework for drawing out and evaluating coherent values.
2026 Operationalising LLM as a Judge with Safety Guardrails. A talk on using language models to judge other outputs, and the safety guardrails needed to do that in practice.
2025-2026 Evaluating Prompt Injection Resistance in LLM Chatbots. Designed the master's capstone project on strenghening LLM's against prompt injection attacks for UTS Transdisciplinary School Masters of Data Science Program for two semesters.
2025 Responsible AI in Practice: Embedding Ethics and Transparency in AI. A talk on where Australia stands on the responsible AI front, and steps we can collectively take to keep up with an ever-evolving AI capability landscape as part of the National AI Week.
2025 AI Made Easy: A Human-Centric Perspective. An overview of human-centric AI, and what it means to keep people at the centre of how AI systems are designed and used.
2022 Learning From Imbalanced Data. A session on how to learn from datasets where one class is much rarer than the others, as part of a Women in AI paper reading club.
Currently reading The Gold-Bug by Edgar Allan Poe.
Next on my list The Power of the Powerless by Václav Havel.
PS: I am a slow reader.
Passing thoughts
I care a lot about x-risks, but I take a passive role in any P(doom) discussions.
I think if we put more weight on resolving the risks we currently have than getting overly hung up on making predictions about the worst case future doom scenarios, we can naturally bring the p of the doom down quite a lot.
That is not to say we should not form ideas of how the future could look like. I only have a problem with it when it overpowers any AI risk related debate and takes the focus away from what needs to be done immediately and in near future, than what's potentially needed in fifty years time.