Healthcare and research data analyst. M.S. Computer Science, Mississippi State University. Currently a Data Analytics Intern in the Agricultural Microbiomes Lab at MSU.
I work on the unglamorous layer analysis actually rests on — validated pipelines, reproducible results, and checks that fail loudly instead of quietly. Every project below states what it found, how it was verified, and where it fell short.
Work-authorized through June 2027 on F-1 OPT, STEM-extension eligible to 2029. Open to relocation. Available now.
nano16s — Oxford Nanopore full-length 16S rRNA taxonomic profiling. The tool belongs to the Agricultural Microbiomes Lab; I authored all 30 merged pull requests that took it from a cluster-bound research script to a documented tool that installs and runs on an ordinary laptop:
- continuous integration, packaging, and a per-run performance report
- cross-platform support — Linux, macOS, WSL2, managed Windows
- failure diagnostics that name the real cause (host clock drift, an empty barcode, a length window that does not fit the amplicon) instead of the misleading error
- data-correctness fixes: the right abundance table, stray files no longer become samples, a failed database rebuild no longer destroys a working one
- a user guide written from an unprepared machine forward, plus contributor docs and linting
clinical-quality-measures — HEDIS / CMS eCQM engine over a 13,810-patient synthetic EHR (820,993 encounters, 10.8M observations). Measures are data, not code: each is a YAML spec with its SNOMED CT and LOINC value sets held separately, so an annual steward update is a reviewable diff rather than a code change with regression risk.
denials-analytics — Three years of CMS Transparency in Coverage filings: 15,545 plan-years, 126.9M denied claims. Only 0.168% are ever appealed, and 40.1% of appeals are overturned. The bottleneck is capacity to file, not claim merit.
ed-throughput-analytics — 16,025 real CDC/NCHS emergency department visits, survey-weighted to 155,397,747 national visits. The pipeline refuses to run unless it reproduces CDC's published total. Between-hospital variation in door-to-provider time is 3.9× larger than between-hour variation — it isn't when, it's where. The peak-hour effect I expected did not survive the weighting, and I reported that as a null result.
readmission-risk — 30-day readmission over 99,343 real inpatient encounters. AUC is 0.664, and the point of the project is that AUC is the wrong thing to judge it on: working the top 20% of the risk list catches 37.7% of all readmissions, at 1.89× enrichment and a number-needed-to-screen of 4.8.
experimentation-toolkit — Five ways A/B tests quietly fail, each measured against known ground truth rather than asserted. Peeking ten times at a null experiment drives the false positive rate to 18.5% against a nominal 5%. CUPED at ρ=0.7 is worth 1.96× the traffic at no cost.
Python SQL R · pandas, NumPy, SciPy, scikit-learn · PostgreSQL, MySQL, SQLite
Power BI, Tableau · Git, CI · SNOMED CT, LOINC, ICD-10, HEDIS/eCQM · Linux, WSL2