Computational psycholinguist. I work on how language-model probabilities relate to human interpretation, using parallel English and Mandarin materials and Bayesian hierarchical models.

• enzh-quantifier-eval: parallel English and Mandarin items for context-forced scalar implicature, scope, numeral and distributivity readings; surprisal-based scoring of base language models, length diagnostics, Bayesian hierarchical analysis.
• minimal-pairs: minimal-pair evaluation that separates ranking from judging (forced choice, pair consistency, critical-word scoring) with paired tests; Kneser-Ney baselines on BLiMP.
• rsa-hope-wh: Rational Speech Act models of a listener who infers who is speaking, with a corpus-informed native prior and an identifiability analysis.

• rag-docs-assistant: retrieval with cited answers over API documentation; BM25, TF-IDF and embedding backends compared on a small benchmark with intervals and paired tests. I write API documentation professionally, and this is the retrieval side of that work.
• morphosyntax-only-parsing: a graph-based dependency parser written from scratch in PyTorch (course project) that asks how much structure gold morphosyntax recovers in Czech and English.

Python, PyTorch, Hugging Face Transformers, PyMC and ArviZ, pandas, scikit-learn. English and Mandarin Chinese.