My independent research explores language model efficiency and information retrieval, with a broader interest in making AI useful and accessible. I develop and evaluate these ideas through practical experiments, then publish the methods and findings.
Find me on arXivAn exploratory ablation study examining how attention, recurrent state, and feed-forward design contribute to a small hybrid language model. The findings are specific to the tested implementation and a single training seed.
A pre-registered study of nested byte-level vocabularies, examining the trade-offs between deployment savings, shared training, and robustness to typographical noise.
A 150-million-parameter language model combining short convolutions with attention, designed specifically for CPU inference. The release includes open code, model weights, and the measurement artefacts behind the evaluation.
A browser extension providing contextual explanations of technical acronyms through a curated dictionary and language models. Evaluated with 25 participants to examine its usefulness for readers with limited digital literacy.
A document segmentation method that uses predicted user questions and dynamic programming to optimise chunk boundaries for search and retrieval-augmented generation.
These papers are arXiv preprints. Code is linked where it’s available.