william garrow

Writing

Notes from the work

Long-form writeups of the problems behind the projects. The first pieces are in the notebook now; each publishes here and, in shorter form, on LinkedIn.

In the notebook

In progress

Research

When the decoder lied too

I built a pipeline to audit a synthetic stand-in. Then a cross-check found my own reader inventing rows for empty tables, inside the frozen result. What I did next, and why the correction is the point.

Research

The reading-level floor: why models can't write at 5th grade

Ask a language model for 5th-grade prose and it settles near 8th. Measured evidence from MedLit, a matching published finding, and what changed when the model did.

Healthcare ML

Your synthetic data is lying to you

A physician caught my demo citing the wrong drug. The pipeline was right; the synthetic record was wrong. What that means for every FHIR project that trusts its test data.

Engineering

Reverse-engineering MATLAB's undocumented table format

How MCOS hides tables inside HDF5, what it took to decode them, and why the fix belongs to everyone. The full story behind mat73-reader.

Until these land here, the work itself is the writing: the case studies carry the stories and the repos carry the details.

Follow on LinkedIn