Ganesh Prasad Shivakumar Email & Phone Number

Seattle, Wa, Us

North Reading, Ma, Us

Collaborating with Applied Scientists & Senior SDEs in building ML solutions to solve Match and Perception problems at Amazon Robotics.

Seattle, Wa, Us

Business Data Technologies - Cairns Team in Seattle- Optimized efficient data retrieval of a Quadrillion bytes of data from Data Planes (Cairns, Andes) by redesigning paginationKey while working alongside developers responsible for orchestrating data storage.- Redesigned the whole serialization framework for paginationKey, making this flexibly adoptable for schema evolutions.- Achieved 30% reduction in response time by devising options to migrate from non-orthodox string serialization methods towards Kryo.- Utilized easily evolutionized paginationKey and helped in versioning paginationKey to push down in-predicate changes.- Accomplished 50% reduction in memory usage for querying shards by using the paginationKey to transform in-predicate query processing from in-memory to a more logical way of processing.- Gained recognition and was featured on the “Employee Spotlight” for devising alternatives to deploy delicate changes without disturbing ongoing traffic using the Amazon Weblab experimentation tool.

North Reading, Ma, Us

Packaging and Allocation Team- Streamlined storage and retrieval workflows and orchestrated efficient methods for data concerning package induction and movement. - Surpassed expectations by achieving millisecond-level lag and established a pub-sub architecture to swiftly share data snapshots among downstream services.- Pioneered a new pipeline using Kinesis Data Firehose and Lambda to detect changes in package information and update snapshots accordingly.- Developed a new CDK package for downstream services, enabling in-memory snapshots that subscribe to the central storage system for real-time updates on package information.- Implemented an in-memory snapshot solution to significantly reduce storage overhead on the service while ensuring updated information delivery in milliseconds through SNS topics.- Achieved 99% synchronization rate between real-world package movement and our system through this transformation, which helped the project garner widespread recognition across the organization.

Boston, Massachusetts, Us

Working as a Teaching assistant for Algorithm Course CS5800 under Prof. Virgil Pavlu.> Responsibilities: Providing solutions for the problem sets, conducting office hours to clarify student's doubts, grading, and post related documents or articles that will help students to ace the course.

Boston, Massachusetts, Us

• Preprocessed US politican twitter data from last 10 years and found top 20 topics using LDA.These processed data is used for determining honesty components using keywords and cosine similarity.• Used FastText embeddings and built attention models to analyze the honesty and trustiwothiness based on the belief and fact speaking topics extracted from topic modelling of twitter data. Able to successfully implement the paper: https://www.nature.com/articles/s41562-023-01691• Led the development of a multi-modal variational autoencoders (VAE) with convnets, transformers and LLMs (BERT, GPT2) effectively calculating visual semantic embedding in political social media (Instagram) data with 400,000 records• Facilitated disentangled representation learning and discovered generative factors by implementing innovative latent space exploration, partitioning it into private and shared spaces, enabling controlled robust image and text reconstruction.• Deployed the model on GCP with advanced loss functions akin to beta-TCVAE, achieving remarkable generative capabilities and results in generating images from text and vice versa with 30% boost in processing, enhancing multi-modal data analysis workflows

Boston, Massachusetts, Us

Working with team of TAs (graduate and PhD) along with Prof. Virgil Pavlu to grade and teaching CS6140 Machine Learning Algorithms:- Facilitate understanding of intricate machine learning concepts, conduct workshops on algorithms such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), hold office hours for clarifications, and provide support with assignments and projects involving advanced techniques like gradient boosting and deep learning.- Evaluate student work, providing constructive feedback on the application of complex algorithms like support vector machines (SVMs) or ensemble methods, and assist in grading assignments, exams, and projects that may involve sophisticated machine learning models aligned with course objectives and guidelines.

San Mateo, California, Us

Designed and contributed to below applications:• Cerebro: Neural network which calculates popularity, user behavior and recommendation scores using analytics data. ( Stack: KubeFlow pipeline, Python 3, TensorFlow, Kubernetes )- Created a customer propensity scoring Learn To Rank models utilizing (ML) Gradient Boosting Decision Trees (xgboost) to rank products based on user analytics, which enhanced Order conversion rates by 20% across users.- Developed Hybrid Memory-Based filtering Recommender Systems incorporating Transfer Learning, Graph-based, and Deep Clustering models utilizing BERT, resulted in 140% enhancement in MAP@5 for product recommendations.• Catalog Pipeline: Built and maintained the catalog indexing pipeline, which is one of the core services for Unbxd platform. (Stack: Argo Workflows, Java 11)• Product Store: Built a service which acts as a source of truth for the catalog of any e-commerce sites, used for differential indexing. (Stack: Java 11, dGraph, AeroSpike, Travis, Kubernetes)• Viper: Auto-Scalable Platform to fetch and transfer catalog from one e-commerce platform to other. Workflow is built through Argo where the Kube pods will be roll up as per need. (Stack: Python3, Django, Celery, Argo, Redis )• Reaper: JAVA Application that validate JSON schema, which is extensively used in catalog pipeline. This application is built on top of schema.org standards. (Stack: Java 11, Redis, SpringBoot, Travis, Kubernetes)• Catalyst: Service that helps in pushing affinity for boosting products in while querying, which is extensively used for AutoParts domain e-commerce websites. (Stack: Java, Redis, Maven deployed using Kubernetes)• Named Entity Recognition: Built neural network models which understands the search query to pop up relevant search results by using the trained models from analytics data. (Stack: Python 3, TensorFlow, Kubernetes)

San Mateo, California, Us

In this period worked extensively on pipelining catalog insertion and inventory management tool.- Application that validate JSON schema, which is extensively used in catalog pipeline. This application is built on top of schema.org standards. This tool helped in detecting the 100% any irregularity in the data irrespective the size of the catalog.Stack: Java11, Pippo, Lombok, Maven, Docker, Kubernetes.- Reduced response time by 25% by deploying an embedding-based item recommendation model trained from Factorization Machine, and serving it as an API on AWS EC2 using FastAPI and SCANN.- Online code editor which enables the customers to add rules to enrich, merchandize and boost products based on the season.Stack: Python3, Django, ReactJS, Docker, Kubernetes.

San Mateo, California, Us

In this period, I worked extensively on Python and variants of Javascript like ES6, ReactJS and NodeJS.- Designed and built Javascript SDKs in Builder design pattern for integration of Search, TypeAhead, Analytics and Recommendations. This helped in reducing platform onboarding time by nearly 60%.Stack: React JS, WebPack, Javascript ( ES 6 ), jQuery, HandleBars.- Designed and built plugins which is capable of integrating search, typeahead and analytics tools to the e-commerce platforms like Shopify, BigCommerce, Magento and Miva. This increased the numbers of SMBs onboarding by 40%, which had a major impact in gaining new customers.Stack: Python3 (Shopify & BigCommerce), Java (Magento-2), PHP(Magento-1), Django, PostgreSQL, ReactJS, Kubernetes (Containerised Deployment).- Backend Search integration tool which helps in integrating with e-commerce search engines like Oracle Endeca Commerce. Through this tool, we were able to bypass the search requests and override/update the response, which helped in search integration without much of hustle.Stack: NodeJS, WebPack, Docker, Kubernetes.

Introduction To Deep Learning

Master Of Science - Ms, Computer Science

Bachelor Of Engineering - Be, Electrical, Electronics And Communications Engineering