May 2019 – Aug 2019 · 3 mos · Vancouver, British Columbia, Canada

• Model-based Reinforcement-learning for Industrial Process Control -- Worked on model-based RL algorithms for learning control in industrial process applications (e.g. manufacturing plants). We investigated techniques for learning robust deep-learning models of the environment dynamics that are accurate in long-horizon predictions and could be used to learn control policies in the offline-RL setting.

May 2018 – Aug 2018 · 3 mos · San Francisco Bay Area

• Imitation Learning in partially observable Markov decision processes (POMDPs) - We investigated deep-learning architectures and algorithms for imitation learning (from expert demonstrations) in POMDP environments using adversarial methods. Please refer to the paper for details - https://arxiv.org/abs/1906.09510

Jun 2015 – Sep 2015 · 3 mos · Bellevue, Washington

• Worked with the GPU modeling team on hardware techniques to make the GPU more amenable to running complex, irregular codes as opposed to only regular, embarrassingly parallel applications. We proposed modifications to the way threads are handled inside a GPU by augmenting the command-processor.
• Implemented new hardware modules in the AMD GEM5 simulator to evaluate the potential benefits on OpenCL applications under the HSA (Heterogeneous System Architecture) platform.

Aug 2011 – Aug 2013 · 2 yrs · Bangalore, India

• I had the opportunity to contribute to the first two iterations of the Intel Xeon Phi cards, aimed at the high performance computing (HPC) market. Our team handled the performance analysis and modeling of scientific and numeric codes (like Lattice Boltzmann, Molecular Dynamics, Weather Forecasting) on the Knights Corner silicon, and on the Knights Landing performance simulator. Of particular interest were improving the parallelization efficiency of the codes on the many-core chips using compiler and programming-model techniques, and also alleviating hardware-level scalability bottlenecks such as contention at shared resources (e.g. DRAM buffers).

May 2010 – Jul 2010 · 2 mos · Fredericton, Canada

• Selected for the MITACS Globalink Research Internship program (https://www.mitacs.ca/en/programs/globalink)
• Worked with Prof. Lichen Chang in the Sustainable Power Research Group on the design of a MOSFET-based amplifier board to handle the dynamic load from a wind turbine generator prototype.