Engineering leadership · Machine learning · Healthcare

I make complex systems comprehensible.

I lead engineering teams and build machine learning systems that ground their claims in sources you can check. My research sits where cognitive science meets software: measuring comprehension, then engineering for it.

See the work Get in touch

Why it matters

Working alone and unfunded in a research course, I converged on the same grounded architecture the industry landed on for explaining medical records to patients. Now I’m looking for the problems the giants won’t solve.

Read the MedLit case study →

Selected work

Work that ships with its evidence

All projects →

MedLit

Takes a patient's FHIR record and returns plain-language explanations at a chosen reading level, each with its citations. Reference ranges decide the lab results; the model is not in that code path.

Converged on the architecture the industry landed on

FHIRRAGFastAPINext.jsDocker

mat73-reader

Reads the MATLAB table objects that scipy and mat73 cannot: a reverse-engineered decoder for the undocumented MCOS format, with a CLI for inspection and conversion.

799 tables decoded from a 3.8 GB research dataset

PythonHDF5pandasPyPI

Cognitive load from eye movements

Classifies overload from eye movements, grounded in Cognitive Load Theory and rebuilt on the real COLET corpus after the synthetic stand-in's numbers came apart. The twelve-coefficient model runs in the page.

0.783 macro F1 on the real corpus, participants held out

scikit-learnEye trackingCOLETResearch

The standard

Every system I ship can show where its claims came from.

On MedLit that means citations keyed by the patient’s own codes and lab verdicts from reference ranges, with no model in that path. On the cognitive load work it meant publishing a correction beside the number it replaced.

Claims with receipts

01234567890123456789.012345678901234567890123456789012345678901234567890123456789

macro F1

cognitive load classifier on the real corpus, participants held out

012345678901234567890123456789012345678901234567890123456789

tables decoded

by mat73-reader, from tables scipy cannot open

01234567890123456789

grounded sources

behind every MedLit explanation

01234567890123456789

LLM calls

in MedLit lab interpretation

Building something at this intersection?

I’m always up for a conversation about machine learning in healthcare, cognitive science research, engineering leadership, or the startup scene in New England.

Start a conversation