Work

Winning project at the Y Combinator Bio/AI Hackathon. Tri-modal immunotherapy-readiness system fusing H&E whole-slide image pathology, bulk RNA-seq transcriptomics, and clinical data to classify tumor microenvironments as Hot/Warm/Cold and support treatment sequencing, clinical trial matching, and agentic co-pilot queries.

Mechanistic interpretability analysis of DeepVariant, Google's deep learning CNN for genomic variant calling. Using sparse autoencoders to profile and cluster model embeddings to decode how representations capture GIAB genomic stratifications, ENCODE functional annotations, human ancestry (1000 Genomes), and cross-species generalization.

Unified analysis pipelines for pooled optical screens at the Innovative Genomics Institute. Includes multi-channel fluorescence alignment, automated cell segmentation, and spatial transcriptomics integration across large-scale datasets.

End-to-end representation learning pipeline using Doc2Vec to encode longitudinal ICD-10 claim sequences into 128-dimensional patient trajectory embeddings across 447,222 MS patients. Applied stratified GMMs to identify prodromal phenotype clusters, validated against Time-to-Diagnosis outcomes.

Bioinformatics pipeline for Cellanome platform data processing, CITE-seq integration, and cell-cell interaction analysis to identify therapeutic targets that rescue dendritic cell priming of T cells in immunosuppressive tumor environments. Won 1st place at UCSF ImmunoX Hackathon.

Local-first tool for analyzing your iMessage chat history. Extracts and visualizes conversation patterns, sentiment trends, response times, and word frequencies across contacts — all without leaving your machine.