👨🏻‍💻 Advitya Gemawat Email & Phone Number

Redmond, Washington, Us

Evaluation on Azure AI Foundry

Redmond, Washington, Us

Responsible AI for Computer Vision

Redmond, Washington, Us

Responsible AI team under Azure ML & Azure AI Studio product lines | Azure AI Platform org

Redmond, Washington, Us

Microsoft AI Development Acceleration Program (MAIDAP) [Program’s acceptance rate ~= 0.2%]Spring 2023 Rotation: Azure Machine Learning (AML) team, AI Platform org & Microsoft Research (MSR)Fall 2022 Rotation: AIOps Gandalf Team, Health & Standards org within Azure Edge & Platform (AEP)Spring 2022 Rotation: Gray Systems Lab (GSL), Azure Data orgFall 2021 Rotation: Signal Quality team, Azure Quality Infrastructure (AQI) orgOne of the most elite New Grad programs in all of Microsoft, located at the "most innovative square mile on the planet"!L60: Mar 2022 - PresentL59: Jul 2021 - Feb 2022 • 7 mos

Bellevue, Wa, Us

Industry Mentor for the 'GenAI Evaluation Copilot' Grad Capstone Project, sponsored by Microsoft.

La Jolla, California, Us

Ideated 'GraphGem: Optimized Scalable System for Graph Convolutional Networks' (Top 6 @ ACM SIGMOD SRC 2021) and advised by Prof Dr. Arun Kumar.Poster Link - https://cs.uwaterloo.ca/~xihe/srcposter2021/Posters/UG1-4.pdfAbstract Link - https://dl.acm.org/doi/abs/10.1145/3448016.3450573

La Jolla, California, Us

An undergraduate Guest Speaker in a Graduate-level Databases seminar course.Gave a talk on Project 'Vista', recipient of the HDSI Scholarship, HDSI Scholarship Project Award, and the CRA Outstanding Researcher Award nomination (Microsoft Research sponsored).Abstract:Scalable systems for ML are largely siloed into dataflow systems for structured data and DL systems for unstructured data. This gap has left workloads that jointly analyze both forms of data with poor systems support, leading to both low system efficiency and grunt work. We plan to bridge this gap for a class of such workloads: feature transfer from Deep NNs for analyzing unstructured (images, text etc) along with structured data. Vista is a new data system that resolves systems issues by elevating entire transfer learning workloads to a declarative level on top of Parallel Dataflow (PD) and DL systems. Vista automatically optimizes configuration and execution of this workload to reduce computational redundancy and improve workload reliability. In this talk, I talk about some of the extensions I worked on for Vista, along with some of my current efforts to expand Vista to support general transfer learning workloads, along with the broader goal to integrate Vista with Cerebro.Speaker bio:Advitya Gemawat is a fourth-year undergraduate student advised by Prof. Arun Kumar at UC San Diego. His research interests lie along the intersections of Machine Learning and scalable Systems, an emerging area which is increasingly referred to as Systems for ML. He's a recipient of the HDSI Undergraduate Scholarship, Siemens SRC Scholarship, and has been nominated for the CRA Outstanding Undergraduate Research Award (sponsored by Microsoft Research). He's previously interned at Fidelity International, PwC, HP, and recently at VMware, and has spearheaded hyperparameter optimization and AutoML capabilities in the Apache MADlib library's Deep Learning module, all of which have been released in the version 1.18.0.

La Jolla, California, Us

Among 20 proposals selected from UC San Diego for HDSI's Undergraduate Research Scholarship Program for the 2020 budget year.• Integrated support and declarative hyperparameter tuning for 5 downstream ML models into the Lab’s CNN workflow optimizer with PySpark.• Evaluated, visualized and compared end-to-end runtimes of 6 downstream models by running experiments with Spark clusters under a distributed set-up.• Automated run-time experiments by writing custom Shell scripts to initialize and run all processes in background, saving 5+ hours per manual experimental run.• Systemized a gap in scalable feature transfer for multimodal analytics by identifying 4500x memory blow-ups with BERT inference workloads.Sponsored by VMware, NSF and NIH

San Diego, California, Us

*Secured the HDSI Undergraduate Scholarship Project Award as part of the graduating class of 2021*Awarded $2500 scholarship as tuition reduction for 2020 to support my educational activities at UC San Diego by contributing to an 'End-to-End Declarative Transfer Learning System for Multimodal Analytics with Deep Neural Networks'.Project 'Vista': An End-to-End Declarative Transfer Learning System for Multimodal Analytics with Deep Neural Networks Mentor: Prof Dr. Arun KumarScalable systems for ML are largely siloed into dataflow systems for structured data and Deep Learning (DL) systems for unstructured data. This gap has left workloads that jointly analyze both forms of data with poor systems support, leading to both low system efficiency and grunt work. We plan to bridge this gap for a class of such workloads: feature transfer from Deep NNs for analyzing unstructured (images, text etc) along with structured data. Vista is a new data system that resolves systems issues by elevating entire transfer learning workloads to a declarative level on top of Parallel Dataflow (PD) and DL systems. Vista automatically optimizes configuration and execution of such workloads to reduce computational redundancy and improve workload reliability. Initial experimental results show that Vista can enable speedups up to 10x compared to existing baseline approaches.We base Vista on 3 design decisions: Declarativity to simplify specification, Execution Optimization to reduce runtimes, and Memory Management and Configuration Optimization to avoid crashes.

Wilmington, Delaware, Us

One among ~56 Apache MADlib contributors worldwide at the time. One of the youngest Apache MADlib contributors in the world. Deep Learning module: Hyperparameter Optimization for Model SelectionAdded support for the following UDFs: ➡️ `generate_model_configs` to perform grid and random search for deep learning model selection ➡️ `hyperband_schedule` utility function for users to pick and choose the schedule they wish to use for hyperband➡️ `madlib_keras_automl` with AutoML capabilities such as Hyperband and Hyperopt with novel algorithmic optimizations for Massively Parallel Processing (MPP) databasesAll contributions officially released as features in MADlib version 1.18.0!Contributions made as part of internship at VMware.

Palo Alto, Ca, Us

Supposedly the first undergraduate student researcher in VMware's history to first-author a research blog on the company's product blog site ["Massively Parallel Automated Model Building for Deep Learning" | media attached below]Advanced Analytics team for Greenplum Database R&D | Modern Applications Platform Business Unit (MAPBU) | VMware Tanzu portfolio | 15 weeks full-time internship• Introduced 5 hyperparameter search algorithms and 3 new APIs with python and postgres to automate deep learning model selection in Apache MADlib.• Pioneered 2 AutoML capabilities with Hyperband and Hyperopt optimized for MPP and Model-Hopper Parallelism to accelerate end-to-end model selection.• Revolutionized hyperparameter optimization MADlib’s DL module in released version 1.18.0, becoming one of the youngest Apache MADlib contributors in the world!

Palo Alto, Ca, Us

Exceptionally selected in HP's MS/PhD Data Science Intern program as an undergraduate. Global Pricing Analytics team, Chief Commercial Office Organization• Formulated a pre-processing automation tool to flag warnings related to bad feature combinations for 50+ pricing models refreshed quarterly for all of HP’s products globally.• Upgraded pre-processing functionality to parse features via 10+ declarative config options to provide warnings and time-series plots.• Accelerated execution speed-ups from O(N^2) to O(N) by integrating on-demand capabilities with the tool.• Amalgamated 4 internal metrics to measure impact of individual feature drivers around ad-spend, price, seasonality, inventory etc.

Contributed to a specialized Web Analytics platform to extract and predict insights related with demographics, user engagements and social ideologies from social media sites using language models and IBM Watson.

Joined the student-run projects group as the first Data Scientist.• Mined 42000+ tweets with #dillydilly from December 2017 to April 2018 using tweepy to access the Twitter Application Program Interface (API).• Extracted qualitative understanding about dillydilly’s usage in contexts of sports, social events, top users, public holidays etc as a whole and as 24-hour time windows through a Topic Model.• Visualized the word embeddings extracted with Stanford's Global Vectors for Word Representation (GloVe) Model using t-distributed Stochastic Neighbor Embedding (t-SNE) and trained a seq2seq Recurrent Neural Network (LSTM) model on embedded tweet data to autocomplete tweets.• Automated detection of token-oriented complaints in Reddit's Comments Data using the Valence Aware Dictionary for sEntiment Reasoner (VADER) model’s polarity and feature parameters.• Engineered a ‘Controversiality Detector’ with feature extraction and stacking Decision Tree based models on a 2nd-level eXtreme Gradient Boosting (XGBoost) Classifier on Reddit’s Comments Data to enhance accuracy by 62.7%. • Predicted the Amazon Q/A data's close-ended answers by 99.726% with a Support Vector Machine (SVM) Classifier.• Developed a Chatbot with Amazon Q/A dataset and text2query mappings and generated most similar questions to user input by leveraging Term Frequency - Inverse Document Frequency (tf-idf) vectors and Semantic Parsing.• Designed a 'question2SQL' framework in a seq2seq NMT model with Keras embeddings and accurately evaluated the position of a query keyword by 80%.

Gb

LoS: Advisory Sub-SBU/Team: Analytics• Designed an ER diagram to represent the data model stored after ETL processes on disparate city-wide government data.• Optimized topic detection to analyze customer reviews with POS-tagging and integrated NLTK’s similarity dictionary to reduce redundancy in category representations by 54.26%.• Fabricated the functionality and UI of a Chatbot supporting autocorrect, multi-language translations, transliteration, feedback using HTML, CSS, Ajax, jQuery and Flask as part of a proposal for a $200,000 project.

London, London, Gb

• Invented a novel tool to automate extraction and classification of specific information within US Companies' Filings by 100%. • Spearheaded the construction of systems and pipelines involving Web Scraping, Information Retrieval and Natural Language Processing (NLP) for investment research purposes.• Achieved accurate topic forecasting of corpora of financial texts by 70% using a Multinomial Naive Bayes Classifier.

Žilina, Sk

Went for ZAMUN and Foreign Exchange to Slovakia and stayed with a host family. Secured the Best Distinguished Delegate award as Delegate of Syria in UN Human Rights Council, discussing measures on eradicating Law Enforcement Corruption in regions with Human Trafficking.

Data Science

Data Science