Deepanshu Mody

Developed and implemented custom registers and instructions for optimizing deep learning algorithms on the RISC-V architecture using the LLVM compiler. Added support for custom vector registers and load and store instructions to significantly improve performance. Streamlined deep learning model debugging by developing Python scripts to modify network architecture, reducing debugging time by 80%.

Achieved state-of-the-art performance on a 3D reconstruction problem, with sensitivity, specificity, and accuracy values within 5% of the state-of-the-art leveraging cutting-edge GNN methods such as GCN, GraphSage, and GAT. Generated a comprehensive dataset of over one million graphs that represented a segment of a molecule as a graph, using raw 3D molecule data. Identified gaps in the field by conducting a literature survey of >100 papers and addressed them by leveraging GNNs.

Developed an OR Tools model in a team of 5 to minimize transportation costs using Google OR-tools constraint solver and devised 4 new constraints to find the optimal path for a fleet of vehicles.

Relevant Coursework: Machine Learning; Deep Learning; Artificial Intelligence; Database Management Systems; Data Structures and Algorithms; Operating Systems; Computer Networks; Discrete Mathematics.

Exchange Semester for Bachelor’s Thesis - Deep learning to predict 3D coordinates.