Selected as 1 of 18 students globally for a competitive research program; Developed satellite image segmentation models using U-Net and Attention U-Net in PyTorch and NumPy.
Trained and evaluated models on ~10k images from SpaceNet and Massachusetts datasets, achieving an F1 score of 0.73, exceeding common benchmark performance (~0.70).
Media
Entered the program with no prior deep learning experience; ramped from fundamentals to implementing U-Net architectures within weeks.