Publications

I started this because I wanted to know how far small local language models could go when answering questions in natural language, like “How much has my grocery spending increased this month compared with the same month last year?” The interesting question wasn't simply whether a model could answer, but how to evaluate a system that had to be fast, small, private, and reliable enough to use on a personal device.

I became interested in whether an LLM's behavior could change simply because I changed the language or context in which I asked the question. Working on the evaluation pipeline made me think much more carefully about how to design evaluation pipelines and execute them; you really do have to be very thorough.

This was my first substantial experience working with a real-world health dataset. I worked with NHANES data and built survey-weighted multivariable logistic regression models, which taught me what it actually takes to go from a research question to an interpretable analysis rather than simply fitting a model.

It was also where I learned end-to-end modelling work: how to perform a literature review, clean data, choose and transform variables, run the analysis, and create useful visualisations.

I initially got into evolutionary optimization because the idea of optimising objectives that can be genuinely conflicting seemed interesting to me. I also liked that non-ML approaches, including genetic algorithms, can work surprisingly well for these kinds of problems.

For our undergraduate project, we studied how NSGA-III's Innovized Progress operators behaved under different machine-learning models. One thing I took away from the project was that conclusions depend not only on the model, but also on the comparison design and the metrics used.