About
Making systems make sense
I’m a software development manager who builds machine learning systems on the side that refuse to make unverifiable claims. My graduate work spans machine learning at Georgia Tech (MSCS) and business at Boston University’s Questrom School of Business (MBA); my undergraduate degree is in computer science, summa cum laude.
Everything I build comes back to comprehension. My cognitive science research measures when understanding breaks down. MedLit turns raw medical records into explanations patients can act on, with citations. mat73-reader opened an undocumented binary format to an entire research ecosystem. The day job is leading engineering teams through modernization, which is making systems comprehensible enough to change.
How I work
Evidence over adjectives. Every number on this site comes from a repo, a paper, or a running system you can check.
Systems I ship show where their claims come from. MedLit cites the patient’s own codes and leaves lab verdicts to reference ranges; the cognitive load model runs in its page with all twelve coefficients on the sliders. When the data behind a result turns out to be wrong, the correction goes up beside the original: in August that meant checking my own decoder cell for cell against a second implementation, because numbers that depended on it were already out, and refreezing the result at 0.783.
See the work →Where I’ve built
Four domains, one habit
Biotech and drug discovery
Preclinical screening platforms, genomic and proteomic analysis tooling, and the AWS data infrastructure underneath them. High-throughput science runs on software that cannot lose a decimal.
Healthcare data
HL7 and FHIR integrations across clinics, PHI compliance, and now grounded language systems over clinical records. MedLit is the first of my FHIR projects I could publish; the earlier ones ran inside clinics, under PHI rules.
Education technology
Large-scale learning management systems, LTI integrations, and build-versus-buy prototyping that saved real money. Learning platforms are comprehension engineering at scale.
Enterprise systems
SaaS compliance products, cloud migrations, and legacy modernization. Leading teams through systems old enough to have their own archaeology.
Now
Current work
The cognitive load model is rebuilt on the real corpus its paper had to synthesize around, audited against the stand-in, and refrozen at 0.783 macro F1 after a correction to my own decoder; the write-up is in progress. fhir-code-lint, the validator that came out of MedLit’s wrong-code story, is on PyPI. MedLit itself is a finished research showpiece whose market six larger launches solved in the same spring. Along the way, getting to know the startup communities in Boston and Maine.
See the work →