Scientific Data Systems · ML/AI Engineering · Computational Biology
I engineer reproducible workflows, structured scientific data products, biological knowledge systems, and reviewable AI applications so computational methods remain reliable beyond a one-off analysis.
At Broad, I built a deterministic medicinal-chemistry triage system connecting compound evidence with final human selection. It makes data integration, validation, and scientist-facing software part of the discovery workflow.
Aggregate description of internal work; public code and a live application are not provided. The triage workflow is deterministic, with scientists making the final decisions.
Virtual-screening methods and evaluation →Scientific Data Systems & Public Software
Metadata transformation, RNA-seq quality control, and reproducible analytical workflows, alongside clearly labeled research prototypes.
Showing all 7 software projects.
Turns inconsistent scientific metadata into validated, provenance-aware structured data.
I designed an agentic workflow in which a supervisor coordinates schema inference, parsing, refinement, and validation. Typed contracts, retries, run manifests, and column-level provenance keep LLM-assisted transformations reviewable and suitable for downstream scientific systems.
Makes bulk RNA-seq quality control consistent, reviewable, and accessible before downstream analysis.
I built a modular R Shiny data product that validates count matrices, aligns sample metadata, calculates library and detection metrics, applies variance-aware PCA, and exports review-ready results. Golem structure, CI, and testthat coverage move QC beyond an ad hoc notebook.
Encodes repeatable omics processing so cohort analyses can move from raw reads to inspectable outputs.
I developed modular Nextflow workflow prototypes for RNA-seq processing from read trimming and FastQC/MultiQC through STAR alignment and featureCounts, alongside genome-assembly and single-cell experiments. The emphasis is portable execution, explicit stages, and reproducible handoffs.
Tests whether experimental context improves bioactivity prediction beyond chemical structure alone.
In a collaborative Broad hackathon prototype, we integrated cell type, assay format, and readout metadata with molecular descriptors, then compared biologically informed QSAR strategies for IC50 prediction.
Explores target-specific QSAR modeling and compound prioritization from public bioactivity data.
Applies ChEMBL data, RDKit descriptors, and machine learning to model compound bioactivity against targets such as SARS-CoV 3C-like proteinase.
Combines spatial structure and gene expression to classify cancer-annotated and other tissue spots.
In a six-person University of Michigan course project, I co-implemented, optimized, and trained a three-layer GCN. It reached AUC 0.91 on a stratified spot split. This is not validation of patient-level metastasis prediction; the report identifies class imbalance and external validation as limitations.
Tests whether a time-aware latent model preserves biological progression in single-cell expression data.
Uses a custom KL objective centered on capture time to shape a 10-dimensional latent space across 14,717 cells and 3,815 expression features, with paired UMAPs comparing true and posterior time structure.
Technical Expertise
The stack matters because it makes analytical decisions reproducible, testable, and easier to review across scientific teams.
Collaboration
I am interested in systems where biological context, model evaluation, validation, provenance, and scalable delivery all matter.