Notes, papers, and ongoing inquiry — questions I'm following wherever they lead.
An exploratory study on 1,617 newspaper-reported suicide cases from Bangladesh — a country with no central suicide registry. We test 32 combinations of imputation, clustering, and evaluation to ask which pipeline recovers real structure from severely incomplete data, and which only recovers how the data was collected.
Our ICCIT 2025 paper on normalizing Banglish — Bangla written in English letters — by treating a sentence as a graph of tokens instead of a sequence. The model reaches 92.64% word-level accuracy, ahead of BiLSTM, Transformer, and edit-distance baselines.