Research
Research
I use first-principles computation and machine learning to make difficult measurements quantitative — and to extract reliable structure from complex scientific and clinical data.
Resonant soft-X-ray scattering · I(q)
Featured · Physical Review Letters
Quantitative and bond-traceable resonant X-ray optical tensors of organic molecules
Makes resonant soft-X-ray scattering quantitative — resolving the orientation of individual chemical bonds inside molecular nanostructures, a longstanding gap in the field.
Resonant soft X-ray scattering and NEXAFS spectroscopy are extraordinarily sensitive to how individual chemical bonds are oriented inside a material — the signal varies with photon energy and X-ray polarization. That sensitivity had never been fully usable, because quantitative analysis requires an accurate optical model with both bond and orientation specificity, and no such model existed.
My work builds one. An algorithm parameterizes and refines first-principles density functional theory calculations against angle-resolved absorption measurements, producing an optical tensor that reproduces data across samples with different molecular orientation and crystalline packing. In practice, it enables label-free orientation analysis of specific chemical moieties inside molecular nanostructures — a longstanding gap in quantitative resonant soft X-ray analysis.
What I find most useful about the result is its shape, which recurs across my work:
01
First-principles DFT
02
Dimensionality reduction & clustering
03
Interpretable representation
04
Refinement against experiment
05
Quantitative optical tensors
06
Better models of orientation & nanostructure
A hard first-principles problem is compressed into an interpretable representation, then refined against real measurements until it becomes quantitative and predictive. See Publications for the full record.
Medical & clinical AI
From 2022 to 2025 I applied the same instinct — extracting reliable structure from messy, high-stakes data — to medicine, as a Data Scientist (MAVERIC) and AI Researcher (National Artificial Intelligence Institute) at the U.S. Department of Veterans Affairs in Boston, and with Dr. Marco Zenati's Harvard MRCAS laboratory.
Clinical-trial matching
NLP and transformer systems that match patients to trials from electronic health records and eligibility criteria, including Molecular Consult for the National Precision Oncology Program; presented at AMIA.
Medical imaging
Deep-learning pipelines on DICOM imaging for liver-cancer screening (the PREMIUM trial) and ARDS classification from chest X-rays, with Grad-CAM / SHAP interpretability.
Research impact & trustworthy AI
NLP for bibliometric analysis of VA research, and fairness / bias audits on clinical models; released PyUMLS-Similarity. This work earned the VA's highest federal performance rating.
Surgical safety
With Harvard MRCAS, multimodal models (EKG, audio, transcribed speech) assessing cognitive load and communication in cardiac surgery, including GANs for synthetic physiological data.
Some of this is open source — PyUMLS_Similarity (semantic similarity over the UMLS ontology) and a clinical-trial concept visualizer.
Earlier work — organic electronics & X-ray spectroscopy
Before the optical-tensor work, my research centered on the structure–property relationships of organic semiconductors — the polymers and small molecules used in solar cells, transistors, and related carbon-based technologies. Using X-ray spectroscopy and scattering, I studied how molecular conformation, aggregation, and interfaces govern charge separation and device performance. That body of work is listed under Publications.