I build systems that turn expert judgment into software.
Physician-trained, bioinformatician, and developer. I work where medicine, machine learning, and messy real-world data meet, and I ship tools that close the gap between how experts actually think and how software usually models them.
As a bioinformatics Ph.D. and an M.D., I turn opaque, effortful, and costly expert judgment into reproducible systems: automating nuclear magnetic resonance metabolite identification, benchmarking on-device LLMs to generate patient notes within the real-world computational and privacy constraints of clinics, and translating painters' compositional logic into parametric code.
I build clinical AI that runs inside clinics, on local hardware, and integrated with real workflows. My focus is documentation quality, privacy-preserving inference, and tools that go beyond benchmark assumptions. In 2020, I led an NSF I-Corps project bringing an NLP tool for clinical trials into an MVP through dozens of user interviews. Good technical ideas die when they ignore workflow, so I design against that.
Benchmarking On-Device Language Models as Medical Digital Scribes · First author · Under peer-review, 2026
Small language models running on local hardware can draft primary-care notes that match human quality, with no detectable age or gender bias, and no patient data ever leaving the clinic.
Neuronally Enriched Microvesicle RNAs in Parkinson's · First author · Frontiers in Neuroscience, 2023
A blood draw carries brain-derived RNA signals accurate enough to flag Parkinson's disease with 94% AUC, opening a non-invasive path to earlier diagnosis.
Semi-Automated NMR Pipeline for Environmental Exposures · First author · Pacific Symposium on Biocomputing, 2021
An automated pipeline identified 79 metabolites that a proprietary tool missed in the same smoking dataset, turning a black-box analysis into a reproducible one.
Full publication list on ORCID
Python, R, JavaScript, Three.js. PyTorch for modeling, SciSpaCy and BioClinicalBERT for clinical NLP, ROUGE / BLEU / entity-level evaluation for LLM benchmarking. Comfortable across local inference, HPC, and AWS.
I reverse-engineer painters' compositional logic into parametric 3D code. Same instinct as the clinical work, different output.
- Hopper · light, geometry, and American solitude as parametric scenes.
- Klint & Kandinsky · early-abstraction form languages, ported to WebGL.
- Lepidoptera scales · biological pattern as generative system.
- Full creative portfolio