• Led development of a vision model to identify maximum land area for accessory dwelling unit (ADU) in residential properties from satellite aerial images across the US.
▪ Sourced aerial imagery data from multiple vendors and perform quality/pricing tradeoff analysis on datasets.
▪ Hired and trained data annotators, cross validated labeling jobs to ensure quality falls within 90% accuracy.
▪ Evaluated three image segmentation models on two datasets, and finetuned final model to achieve 90% mIOU and 85% accuracy improvement in predicting max land ADU area, matching human baseline performance.
▪ Identified bottlenecks in training pipeline including third-party library performance issues and improved training speed by 2.5x on multi-GPU training cluster.
• · Developed an interactive front-end using React and Mapbox to visualize image segmentation results as overlays on satellite imagery, enabling users to hover over areas to display physical dimensions of covered land and obstacles.
▪ Implemented dynamic identification of the largest feasible ADU construction zone with an algorithm implemented in JavaScript to calculate and display optimal quadrilateral fit (square/rectangle) for efficient land use.
