cat projects.md
Projects
SeedBench-Bio
Most demos of "AI-assisted code review" are unfalsifiable, and none are tested against the specific methodological standards bioinformatics pipelines actually need — correct QC assumptions, valid statistical tests, no data leakage. The model returns a paragraph of feedback with no way to check if it was actually good, or if a different prompt would have caught more. SeedBench-Bio is a benchmark I built to test that: seeded ground-truth errors into real bioinformatics pipelines, tested against 5 different prompting strategies.
EEG Seizure Detection
Majority walkthroughs of “AI for EEG seizure detection” stop at reproducing a fixed pipeline and reporting one accuracy score. But with 98.4% of windows being non-seizure, that score is misleading, as a model that never detects a seizure still achieves 98% accuracy. Randomly splitting windows from the same patient across train and test can also let the model memorize patient-specific EEG signatures rather than learn general seizure patterns. This project is a small study I ran on the CHB-MIT pediatric EEG dataset to test that directly: I compared three models (logistic regression, random forest, XGBoost) on engineered time-domain, frequency-domain, entropy, and Hjorth-parameter features, then measured how much performance actually dropped under leave-one-patient-out cross-validation versus a naive random split.
ExprCompareR
Comparing RNA-seq and protein expression across tissues normally means manually pulling data from GTEx and the Human Protein Atlas, then hand-rolling normalization and stats. ExprCompareR is an R package that automates the pipeline end-to-end — querying both databases, computing Spearman correlations, flagging RNA/protein outlier genes, and comparing cancer vs. normal tissue expression — all through a Shiny app.
BMP7 Signaling in Medulloblastoma: Bulk RNA-seq Analysis
Recreated and extended the analysis behind Ohata et al. (2025) on BMP7-driven oncogenic signaling in medulloblastoma — cleaning, normalizing, and running differential expression on their bulk RNA-seq dataset (GSE229150), then following up with thresholded and non-thresholded pathway enrichment and building an enrichment map in Cytoscape.
Arabidopsis Gene Regulatory Networks
Newly discovered transcription-factor dimer interactions from dDAP-seq were sitting outside existing plant biology databases, so I parsed narrowPeak files against the TAIR10 genome annotation to extract the AGIs of interacting genes, then built a UI extension to visualize these new protein-DNA and protein-protein interactions in AIV2/ePlant.
SNV Pathogenicity Predictor
Doctors need fast, low-cost ways to interpret Sanger sequencing results for personalized treatment, so our 7-person hackathon team built an ML pipeline that classifies SNPs as pathogenic or benign — combining ClinVar annotations with nucleotide-transformer sequence embeddings to train a random forest classifier. I mined and annotated the genomic datasets (mapping each SNV to its gene, exon, and regulatory feature) for training/validation.
Study Buddy Finder Application
Students struggle to find study partners who share their courses and schedule, so this Java app (built with a 5-person team) matches users by shared courses and availability, falling back to program-level matches when no course overlap exists — all built with Clean Architecture and integrated with the Slotify scheduling API.