Built a BERT-based classifier to extract flood events from 40,000+ tagged Bangladeshi news articles, then used the resulting time-series (validated against Sentinel satellite data) to help the Bangladesh government develop a flood-index insurance product. Presented at AGU; published as a pre-print.
Built a phrase-grounding pipeline using YOLOv3 and BERT to extract and link images and captions from research papers, reaching 85% accuracy, then constructed a searchable knowledge graph from the results. Also classified dosage-response curves from the extracted features, reaching 92.7% accuracy with AdaBoost.
Extracted syllabi and mission statements from college websites to build a quantitative measure of liberal-arts education across institutions, and developed Python scripts to streamline the project's underlying SQL database interactions.
Used Google Vision and OCR to extract pill features and prescription-bottle imprints, then built a multi-dimensional embedding from the collected data to train RandomForest and SVM classifiers for precise pill identification.
Experimentally demonstrated that a "Liter of Light" water bottle outperforms a glass plate at illuminating low-light spaces like slums. Named a Regional Finalist at the Google Science Fair, and published and presented the research in the Journal of Basic and Applied Engineering Research.