<projects>

Achieved 7-8% performance improvement by sampling and averaging multiple actions over baseline multi-agent algorithms

Achieved 25% higher rewards in upscaling predator-prey environments by extending distributional RL to multi-agent settings

Proposed active learning & contextual reward decoupling approach to improve stability in collaborative-competitive games

Built scalable cooperative multiagent systems for Robosoccer environments by implementing FSMs, coordinated plays, etc.

Inspected end-to-end Warehouse Management Solutions by implementing RRT/RRT* Planning Algorithms on real-life robots

Worked on skills like passing & defense on top of C++ framework; Controlled movement using p-controller in ROS Turtlesim

Developed Layer-adaptive memory optim. algorithm with tradeoff in saving 50% GPU memory vs 100% better execution time

Trained neural networks with lower GPU budget by optimizing Extra forward computation & CPU-GPU transfer time in CNNs

Achieved 13% better execution time than CPU-only implementation and implemented VGGNet & ConvNet on our algorithm

Achieved massive improvements of 20% in identifying Wikipedia Editors leaving the platform helping in early retention

Extracted the User information of Wikipedians and performed sentiment analysis (65%acc.) to gauge levels of satisfaction