DocuMint: Docstring Generation for Python using Small Language Models
Report | Code | HuggingFace
Benchmarked small language models on their docstring generation quality, created a fine-tuning dataset using the FOSS ecosystem, and fine-tuned Google’s CodeGemma. Released the dataset and the fine-tuned model.
Artificial Intelligence Assignments
Developed and delivered programming assignments for graduate level AI and ML courses at University of Memphis. Topics include Deep Q Networks, Proximal Policy Optimization and Finetuning of a Transformer.
BarterBaron: A commerce app based on barter system trade
Developed a ebay-like commerce platform using Ruby on Rails framework with features like chat, content search, and user authentication (won “Best project in the class” award).
Robustness to Input Corruptions and Adversarial Examples in Steering Angle Prediction via Self-Supervision
Examined the use of self-supervised learning in steering angle prediction task to enhance model robustness under conditions such as snow, fog, frost, pixelation, and blur.
Latent Representation of Inputs: A Defense Against Adversarial Examples in DQN
Used feature squeezing to improve the performance of a Deep Reinforcement Learning algorithm, DQN trained to play Atari Pong, under adversarial attacks.
Distributed Hyper-paramter tuning of Neural Networks
Parallelized the hyperparameter search objectives in grid search and random search using Distributed Hash Table so that computational resources are shared across multiple nodes to accelerate the search.
Delta Design 3D Printer
Designed and developed a working prototype of a 3D printer using Arduino Mega (Senior year Engineering project).