Build data driven models to forecast SQL, CPU and Blob costs for core services while optimizing APIs and Stored Procedures.Lead gathering and analyzing SQL metrics that allow understanding the capacity consumption and provide forecasting models for both capacity and cogs of steady state on the SQL layer.Identifying areas of optimization opportunities for both storage and CPU consumption patterns on the SQL layer.Disable and Deprecate some legacy services that helps the service stay COGs positive.Work with multiple stakeholders in analyzing the data, implementing config and infra changes in Cost neutral way.

Part of the Core team for One Drive Sharepoint.

• Worked on Elastic Search, Clustering, Step functions, Lambdas and DDBs to create and maintain the Experiment Management System for Sagemaker, which provides a platform for ML engineers to scale, deploy and run ML Experiments on AWS.• Worked on GDPR compliance, Business Intelligence Integration, design and maintenance of multiple features in the Experiment Management Platform of Amazon Sagemaker Studios for Re:Invent 2019• Currently Working on a Region build and developing a cross team ML Pipeline project using Kubernetes, Model Explainability and Elastic Search

Project 1: Human vs Bot behavior and pattern recognition for multiplayer games without a reward policy. Using Kernel Corex,RNN and LSTMsProject 2: Conducted keyword operated topic modelling to extract ranked documents (TF-IDF & COSINE Similarity) on unlabeled languages with minimum domain knowledge such as Oromo, Tigrinya and Tamil using Guided Latent Dirichlet Allocation and Anchored Corex with weak supervision. Improved Identification accuracy by 2% and aided other models such as Chinese Room Translator in the process. Took part in 12 DARPA Evaluations.Scenarios ranged from disasters to military incursions whereas the documents ranged from twitter feeds to articles and the results were compiled and presented at the April 2018 DARPA conference under the paper titled, “Identifying Scenario Relevant Documents for Low Resource Unlabeled Languages”

Develop a new machine learning system model for the company by incorporating Reinforcement Learning with GAN/CNN/Deep Learning techniques to create a Personalized Agent with multiple sub-models.Multi-input analysis combined with multi skill answer generation to create a social as well as task agent which has dialog and conversational skills.Used DialogFlow to create a sample bot to generate text data and launched it on GCP followed by a stateless RNN bot using Tensorflow.Created a multi-platform (text, voice and phone call) agent to generate conversational data which were then fed into a context sensitive seq2seq model to create a fault tolerant agent. Combined the agent with existing bots/personal avatars to see how they function. Increased the hit ratio by about 12%.

This directed research (under Dr Benjamin Nye and Dr Bill Swartout) is focused on expanding a natural language dialog system so that it will structuremore human-like and natural conversations, despite being restricted to a limited set of recordedutterances Research focused on natural language dialog system to mimic human-like natural conversations, despite being restricted to a limited set of recorded utterances using LSTM, word2vec, linear regression and other classifiers, also enforced Reinforcement Learning. Focused on maintaining continuity, anaphora and iterative refinement to diagnose incorrect interactions. These agents mimic army/STEM personnel who lack the freedom of time to attend Career Fairs but can do so vicariously through our agents. Increased the hit coefficient from 48% to 65% and ratings from 3.4* to 4.2* over 5 Evaluations.1. Continuity: Maintaining continuity of conversations and using probabilistic methods to enableanaphora in them.2. Quality: Improving the quality of answer and avoiding repetition. Quality of answer is a subjectiveterm, but it can be evaluated using multiple metrics such as a reward based assessment.3. Iterative Refinement: Exploring ways to diagnose incorrect interactions, such as incorporate auser rating schema.These were the tasks accomplished.

• Worked for a 5 man Startup including myself to create, deliver, update and maintain ChatBots for companies such as:• Denver Broncos, Arizona Cardinals, Sacramento Kings, Olympics Channel, Publix.• Used a combination of NLP and Bot frameworks to deliver the above results with App Insights. First 4 Bots were showcased in F8 Facebook conference by the CEO, Tony Marks.

• During the internship, worked in a 5 man team to explore Bot Framework, IOT Hub and NLP using Rest APIs and LUIS/CRIS.• Developed a support BOT for an internal Microsoft tool called OSE which removed human-human for error handling. We further added multilingual and speech to text. The bot reduced manual error handling by 60%. Azure services, LUIS, Bot Framework along with N grams and Tokenization was used to achieve this task.