Selected work

Clinicians wrote their notes after the fact, from memory, at the end of a shift. The record was thin, late, and occasionally wrong in ways nobody could audit.

A speech pipeline that handles overlapping speakers and clinical vocabulary, feeding a language model constrained to the SOAP structure so the output lands in the fields the record system expects rather than as free prose. Handling was built to HIPAA and GDPR requirements from the first commit, not retro-fitted before launch.

Clinical language is exactly where general transcription falls over, and a summary that invents a symptom is worse than no summary at all. Most of the engineering went into the checks around the model, not the model itself.

Preparation meant buying a book and guessing. Feedback, when it existed, came from a tutor days later and rarely explained the band score.

Question generation across the four papers, automatic scoring models for writing and speaking that return a band with the reasoning attached, vocabulary tools that adapt to what the candidate keeps getting wrong, progress tracking, and a full administrative panel for the people running cohorts.

Scoring is the whole product. A model that marks generously is pleasant and useless, so the scoring was calibrated against real examiner marks before anything shipped to candidates.

Employee records lived in a shared spreadsheet, approvals lived in email threads, and reporting meant somebody rebuilding the same pivot table every month.

Employee records with a proper permission model, approval flows that carry their own audit trail, and reporting built on defined views so that two managers asking the same question get the same number. Delivered as a normal web application their own IT team can host and maintain.

The difficulty in internal tools is never the software. It is that the existing process has undocumented exceptions everyone relies on, and finding those means sitting with the people who do the work.