Diligent

I cleaned and explored board and governance documents in Python with Pandas and NumPy, and built scikit-learn classifiers for regulatory change detection, grading how big a change each incoming regulatory update was, so compliance teams could triage what mattered instead of reading everything. I assembled and cleaned the labelled dataset behind them, ran the evaluation, and built the internal dashboards that showed the team how the models were performing. The data scientists on the team guided the work, and a good share of it was answering what other teams brought over: taking a loose business question and turning it into something the data could actually settle.

Governance material arrives as documents, not as tables. Before a classifier can read a regulatory update, the text has to come out of the document intact. Headings and clause numbering carry meaning and have to survive. Page furniture does not, and if it stays it becomes the loudest signal in the data. That extraction and cleaning was most of the work, as it usually is.

The labels came from compliance experts rather than from the data, which made the labelling scheme the real problem. Two experts reading the same regulatory update would often disagree on whether it was a moderate change or a significant one. Measuring that disagreement mattered more than the model did, because a classifier cannot be more consistent than the labels it learns from.