At Johnson & Johnson, I built a cost-aware pipeline to extract physician information from hospital schedules in PDFs and images. It uses Camelot, OCR, and vision-language models with quality-based fallback, alongside a Streamlit interface for review and annotation.
I also developed models to identify likely early adopters for drug launches using HCP profiles, HCO attributes, and historical sales of analogous drugs. I engineered features from internal ratings and commercial data, then implemented validation in Kedro.
At UT Southwestern Medical Center, I developed a Mask R-CNN pipeline for segmenting nuclei and classifying six cell types in H&E-stained tissue. By adapting the loss function and training strategy for partially annotated samples, I recovered approximately 20% of the training data; the evaluated pipeline achieved an 82.5% detection rate and 82.0% classification accuracy.
At the Donald Danforth Plant Science Center, I developed leaf segmentation and tracking workflows and contributed Python modules, tests, and documentation to PlantCV. My work spans computer vision and machine learning for biomedical imaging, plant phenotyping, and document understanding.

