Performance and Scale - ML

CS 5520 - Mobile Application Development - Under Professor. Dr. Daniel Feinberg

CS 3700 - Under Professor Dr. Alden Jackson

CS 6650 - Building Scalable Distributed Systems under Professor - Dr. Saripalli Prasad

Fundamentals of Cloud Computing

Networks and Distributed Systems (3700)

◦ Conducted research on innovative architectures for deep recurrent multi-modal neural networks using mixed precision training to fit saidnetworks on memory constrained GPUs and utilize Tensor Cores◦ Engineered a highly scalable, multi-cloud capable and efficient SaaS platform delivering exceptional throughput per clientand achieved sub-200 ms end-to-end latency, empowering businesses to leverage near real-time applications and providing a competitiveedge. Employed a variety of cloud-native ideas including containerization, microservices, load-balancing, WAFs, and, middlewares◦ Implemented and automated deployment management policies reducing deployment transition times◦ Championed a streaming-model back-end architecture, achieving drastic throughput increases.◦ Enhanced network security practices, automated SAST & DAST security testing, and trained development and deployment teams◦ Built least permissive privacy-first protocols ensuring only secured access to Personal Identifying Information(PII)◦ Built logging and validation systems to improve visibility of metrics, ensure compliance with the data protection laws, and speed up

◦ Optimized computer vision models to match or surpass SOTA benchmarks while achieving many fold reduction in memory and computerequirements before quantization by implementing a range of compression techniques including pruning, and knowledge distillation◦ Designed and trained model architectures optimized for various target hardware platforms(hardware-model co-design), ranging fromGPUs to DSPs improving performance over base architectures◦ Developed custom parsers and converters between multiple ML training and inference frameworks including Darknet, Tensorflow,PyTorch, ONNX, nGraph(Intel), and TensorRT enabling the use of novel model architectures including Enet.◦ Leveraged parallel multi-stream inference to boost performance on TensorRT and optimized CPU performance by leveraging asynchronous scheduling◦ Built the SDK to operate across Windows, and Linux, platforms running inference on CPU(x86, AVX), GPU(CUDA), and dedicated edgeinference accelerators - ARM NN and Qualcomm SNPE. Recurrent blocks of code were developed in C++ with Python wrappers◦ Established automated build and test pipelines on Bitbucket to ensure stable development and improve developer throughput

My learning responsibilities included learning safety mechanisms followed by the industry, and developing a solution using Augmented reality and Deep Learning to allow for real-time monitoring of the status of machines and internal components for preventive maintenance. I was also deeply involved in creating another proof of concept using Deep Learning to allow for the automated testing of gas-filled parts, thereby reducing possible human error.Tech Stack:Tensorflow, OpenCV, Python, Android Application, Industrial IOT, MQTT, LabelImgThe opportunity to perform the complete machine learning dataset and model development and the subsequent fine-tuning gave me the opportunity to understand the nuanced effects each process in the chain can have on the performance of the final model.

I was involved in : - Multiple domains and roles including but not limited to porting code from MATLAB to c++, accelerating said algorithms using OpenCV and leveraging parallelism, identifying functional criteria and constraints for said algorithms, and identification of suitable hardware for the application of the same. - Developing a proof of concept using Machine Learning and Deep Learning, in specific, to solve the problem of identifying human error and verifying compliance of set criteria.Tech Stack:a. Tensorflowb. Pythonc. OpenCVd. LabelImg