Engineering and applied work — research tooling, evaluation harnesses, and data-science pipelines. I'm building toward enterprise-grade AI applications; for now, here are projects I've shipped.
A visualization pipeline for single-cell transcriptomics embeddings from the Human Reference Atlas (HRA) and the HuBMAP consortium. It progresses from raw UMAP mappings through tuned PCA + t-SNE combinations over ~453k cells, isolating the top 24 cell types for clearer biological interpretation — turning massive single-cell datasets into legible maps.
Probing-based toolkit for analyzing moral reasoning trajectories in LLMs — linear probes for ethical-framework attribution and lightweight activation steering across reasoning steps.
Evaluation harness for LLM belief resistance under strategic persuasion, built on the Source–Message–Channel–Receiver (SMCR) framework across six models and three domains.
Dataset and code studying how ChatGPT rates the quality of natural-language explanations compared with human assessments (LREC-COLING 2024).
A clean replication of efficient LLM fine-tuning with LoRA (low-rank adaptation) — a compact reference for parameter-efficient adaptation.
Dataset supporting the study of ChatGPT vs. human annotators in explaining implicit hate speech (The Web Conference, WWW 2023).