Research.

Connectome datasets like FlyWire and BANC come with annotations: cell types, predicted neurotransmitters, and other labels that people build analyses on top of. I wanted to know how stable those labels actually are, and what an honest answer looks like when the data runs out.

So I built comparisons I could rerun. Each one freezes its cohort, pins the exact release it reads from, and keeps an explicit "unknown" instead of quietly dropping cells that don't have a label. When two releases disagree, the workflow points at where. Every case study was checked again by an independent review before I counted it.

The findings are about the annotations and how reproducible they are. They say nothing about what a circuit does.

There are large public datasets that measure what happens when you change one amino acid in a protein, and there are public structures for many of those proteins. I wanted to see how far the structure alone gets you in explaining the assay results.

I worked through PTEN, NUDT15, TP53 and G6PD. For each one I wrote the analysis plan before looking at the outcome, mapped variants onto the structure, and used rank statistics, permutation tests and negative controls. Some of the results were negative, and those stayed in. What a result means also depends on the assay, so I read each one against its own measurement rather than pooling them.

This is analysis of public data. None of it is a statement about any patient or any diagnosis.

Since March 2026 I've been a student researcher with gait rehabilitation research teams. Part of the work is helping during participant sessions while the Vicon system records. The rest happens afterward: labeling markers, organizing trials, and checking that the data is clean enough to analyze.

One piece I spent real time on was an older MATLAB gait-analysis workflow that had been ported to Python. Getting it right meant chasing unit mismatches, fixing how gait events were detected, and making the code fail loudly instead of returning a plausible wrong number. I checked the repaired version against representative trials before handing it off.

This is my part of a team's research. The clinical questions, and the results, belong to the lab.

This is the same motion-capture question as the gait work, asked in a very different room. The collaboration studies ergonomics in urologic surgery: how surgeons hold their bodies over long procedures, and what that costs them physically.

My part was building video-based motion-tracking workflows and helping capture and analyze the other streams alongside them, including EMG, heart rate and motion sensors. Keeping those streams lined up so they describe the same moment is the same problem CaptureSuite works on.

From September 2025 to April 2026 I worked remotely as a student researcher with the Joshi Laboratory at Georgetown. The work was mostly the plumbing that makes molecular-dynamics research possible: finding trajectories, cleaning them, and putting them into one consistent format.

On top of that I computed features like PCA and RMSF, and wrote checks for AI-generated protein rollouts. RMSD, radius of gyration and bond geometry catch a structure that looks fine at a glance and is physically wrong. I also turned papers from the field into concrete validation plans the group could test against.

I joined the Tsai Lab at CSULB in January 2025 and stayed through September. The lab studies sex differences in brain and behavior.

Most of my time went into scoring behavior from video: automated tracking gave a first pass, and I verified it by hand. I analyzed results in GraphPad Prism and helped with mouse handling, genotyping and tissue collection.