Purushothaman Saravanan

As a Graduate Research Assistant at North Carolina State University, I played a pivotal role in the development and implementation of the NDBlock algorithm. This innovative algorithm excels in efficiently identifying and optimizing block alignments within N-dimensional tensors. Noteworthy achievements and contributions include:

• Demonstrating the algorithm’s efficiency through extensive testing on diverse tensors with varying sparsity rates.
• Incorporating OpenMP for parallelization, resulting in significant improvements in execution times.
• Showcasing remarkable effectiveness, especially for tensors with modes equal to or greater than 5, where the parallel implementation achieved up to a 15x speedup.

As an AI ML Intern at Kodak Alaris in Raleigh, NC, I played a key role in project management. Notable responsibilities and accomplishments include:

• Project manager for automating call center activities, overseeing the development of an automated email response system and an AI-assisted chatbot.
• Demonstrated strong leadership skills by facilitating coordination between project stakeholders and actively participating in project planning.

Part of Compiler Optimization Course work

• Implemented a compiler using Flex and Bison for lexical analysis and parsing, respectively, demonstrating proficiency in parser generator tools
• Developed an LLVM bitcode generator for a custom expression language, showcasing expertise in code generation and LLVM infrastructure

My involvement in this project allowed me to explore the depths of TensorRT and contribute to the community by providing these custom plugins.

I played a crucial role in open-sourcing custom operators by developing plugins for TensorRT. These plugins enhance the capabilities of TensorRT with custom layers that I designed and implemented. Here are some of the notable plugins I contributed:

• CosLU Plugin: Implementing the Cosine Linear Unit (CosLU) activation function for improved neural network architectures.

• Transformed a PyTorch-based model into optimized formats, including TorchScript, ONNX, and TensorRT
• Conducted a comprehensive performance assessment to measure inference speed, with the ONNX runtime and TensorRT execution provider outperforming other methods by delivering a remarkable 35% performance boost compared to the original PyTorch implementation
• Employed NVIDIA Nsight Systems to meticulously profile the various approaches, providing deep insights into CUDA API calls and GPU utilization patterns

• Deployed highly available chatbot solutions on AWS, utilizing both AWS Elastic Kubernetes Service (EKS) and serverless models with AWS Lambda
• Utilized a Flan-UL2 model as a SageMaker endpoint and implemented custom data retrieval tools using DynamoDB to enhance the functionality of the chatbot
• Implemented the agent orchestrator in AWS EKS and incorporated Horizontal Pod Autoscaling to ensure optimal performance under varying loads
• Managed SageMaker endpoints with automation and autoscaling policies through Amazon Controller for Kubernetes ACK

• Developing a robust understanding of data science principles and machine learning techniques to tackle real-world challenges.
• Exploring optimization techniques for enhancing the performance of machine learning models, with a focus on computer vision applications.
• Embarking on the exploration of machine learning compilers, delving into the intersection of software and machine learning infrastructure.

Relevant courses:

• Neural Networks
• Advanced Machine Learning
• Efficient Tensor Computation for AI
• Cloud Computing
• Computer Vision
• Compiler Optimization and Scheduling (Audit)