Umesh Kumar Gattem

3440 John Hinkle Pl · Bloomington, IN 47408 · (812) 272-5756 · umesh.gattem@gmail.com

Greetings, my name is Umesh Kumar, and I am an enthusiastic developer and designer with extensive experience in the field of Artificial Intelligence and Deep Learning technologies. I have worked with various backends such as TensorFlow, Keras, and PyTorch. My proficiency in Python programming spans over seven years. As a highly motivated AI engineer, I have been involved in the design of tools for visually constructing data transformation recipes, developing Deep Learning models, and end-to-end pipelines for Data Scientists.

I am an excellent team player who enjoys contributing to product development by taking proactive initiatives and assuming responsibility for projects. I am excited to apply my skills and expertise to real-world projects that have the potential to positively impact thousands of people and make their lives easier.

Work Experience

Indiana University, Bloomington

Associate Instructor

  • Recently, I served as a Teaching Assistant or Associate Instructor for the Department of Statistics at IUB, specifically for the Applied Statistical Computing Course.
  • The course curriculum covers several topics, including Statistical Computing using R and C/C++, among others.
  • In the past, I served as a Teaching Assistant or Associate Instructor for the Introduction to Statistics Course offered by the Department of Statistics at IUB.
  • The course covered several topics, including Probability Models, Statistical Methods, Maximum Likelihood, Method of Least Squares, and distribution functions, among others.
  • As part of my responsibilities, I conduct weekly office hours to assist and clarify students' doubts and questions related to the course.
  • Additionally, I supervise graduate and PhD students by creating assignment questions, grading their work, providing feedback, and offering solutions to their queries.

August 2023 - May 2023

Razorthink Technologies, India

Artificial Intelligence Engineer

  • I was part of the Razorthink AI Platform Product team, responsible for developing a tool that enables data scientists and analysts to visually construct data transformation recipes, deep learning models, and end-to-end pipelines.
  • In my role, I designed various modeling libraries using cutting-edge technologies such as Transfer Learning, Training and Inferring models, Tensorboard, TFHUB models, and Distributed Training, exploratory data analysis (EDA), and end-to-end pipelines
  • Moreover, I developed the RZTDL Library, a patented deep learning framework that leverages different backends of Tensorflow and Pytorch to support all Deep Learning operations like CNN, RNN, LSTM, Attention, etc.
  • Designed a multi-stage CNN model for analyzing multidimensional time-series data, resulting in a GINI score of 72 forpredicting customer propensity to buy insurance for our esteemed banking clients.
  • Utilized LSTM network to develop a churn prediction model by conducting a comprehensive analysis of demographics and skewed transactional data, resulting in a GINI score of 68 for one of our largest telecommunications clients.
  • Led a team of skilled professionals in developing a Python SDK, which facilitated the generation of blueprints, automated the process of generating code, and created APIs for seamless integration with web applications.

July 2016 - July 2021

Education

Indiana University Bloomington


Master's in Data Science

CGPA: 3.8/4.0

August 2021 - May 2023

Anil Neerukonda Institute of Technology and Science


Bachelor's in Computer Science and Engineering

CGPA: 8.05/10

August 2012 - April 2016

Projects

Facial Recognition on Google Photos    Github   Poster

  • The aim of this project is to construct a personalized image classification system for Google Photos by applying computer vision techniques based on deep learning.
  • Rather than relying on traditional facial recognition algorithms, this project intends to use a custom dataset and implement transfer learning to extract high-level features from pre-trained models like Inception and ResNet.
  • To address the challenge of limited data, the project will employ various data preprocessing techniques such as data augmentation, cropping, and rotation, and may also leverage advanced deep learning models like Generative Adversarial Networks (GAN) or Autoencoders to generate additional data.
  • Ultimately, the goal is to develop a customized model that can accurately classify images within a user's Google Photos collection.
  • The project also seeks to create a comprehensive pipeline for the Google Photos System that can perform person detection and recognition, utilizing the YOLO (You Only Look Once) object detection model to identify faces of individuals by repurposing it from the COCO dataset.
  • The system will be capable of analyzing both individual and group photos, identifying all faces in the image, creating boundary boxes around each person, and correctly labeling each individual with their corresponding name.

Feb 2023 - Present
Empirical Analysis of Multivariate Time Series forecasting using Graph Learning    Github

  • Our research focused on exploring MT-GNN, a cutting-edge model for graph learning and modeling spatio-temporal information. To evaluate its effectiveness, we compared its results with several baseline models we constructed.
  • MT-GNN is a highly complex deep learning model, comprising multiple components that interact with each other. In order to gain a better understanding of the model, we invested significant time in studying it and reviewing the official code that was provided along with the paper.
  • We conducted empirical analysis on multiple datasets, including Traffic Dataset, Solar Energy Dataset, Pems-D7, Paris Mobility, and Energy Consumption, and presented our findings.

Aug 2022 - Dec 2022
Optical Character Recognition using Layout Language Model    Github

  • In this project, my objective was to detect the boundary boxes of text in a given document or image file. To accomplish this, I employed several libraries and compared their respective results.
  • Notably, I utilized technologies such as PaddleOCR, Tesseract OCR, Google API for image recognition, and Layout Language Model.
  • To evaluate the effectiveness of these libraries, I used a dataset from Wantok images, which is a rare language with a complicated format that makes it challenging to predict the text based on the layout and structure of these images. Despite these difficulties, I successfully extracted and recognized the text boundary boxes with satisfactory accuracy.

Aug 2022 - Dec 2022
IPL Dataset Visualization    Github   Slides

  • As a part of my Data Visualization course, I worked on visualizing the IPL Dataset.
  • This dataset contains the statistics of all IPL teams and players across all IPL leagues that have taken place until now.
  • My project involves creating various visualizations, including but not limited to: 'Team performance at home ground', 'Team performance away from home ground', 'Best performance at a non-home ground', 'Strike rates and averages of batsmen', 'Strike rates and averages of bowlers', 'Batsman performance against specific bowlers', 'Bowler performance against particular batsmen', and potentially many more.

Oct 2021 - Dec 2021
Razorthink Deep Learning Framework (RZTDL)

  • RZTDL is a cutting-edge deep learning framework that is patented and developed using various backends such as Tensorflow and Pytorch.
  • It has a wide range of features such as distributed training, transfer learning, Tensorboard, and TF Hub Models.
  • As part of my responsibilities, I was in charge of supporting all Tensorflow operations like Layers, Operators, Metrics, etc.
  • I was also responsible for implementing backend APIs for training and inferring deep learning models.
  • Additionally, I played a key role in the development, testing, and deployment of production solutions.

July 2016 - July 2021
Python Parser and SDK

  • During my time at Razorthink AI, I was involved in various Python SDK projects related to the Razorthink AI product.
  • One of my responsibilities included creating Python libraries that contained parsing logics for generating blueprints for different Tensorflow layers.
  • These blueprints enabled users to easily drag and drop the required blocks, specify parameters and then convert the JSON code to the required Tensorflow code, or vice versa, for training their models.
  • Additionally, I had the opportunity to work on Micro web services using Flask Python and swagger-api-client modules.

July 2016 - July 2021

Skills

Programming Languages & Tools
Technical Skills
  • Frameworks - TensorFlow, PyTorch, Keras, Sci - kit, Pandas, Numpy, Flask, OpenCV, Pyspark, FastAPI, Uvicorn
  • Machine Learning - PCA, T - SNE, TFHub, Neural Networks, Clustering, Transfer Learning, Inferring models
  • Databases - MySQL, Postgres, MongoDB, SQLite
  • Big Data - Spark, Hadoop, Kafka, S3, Horovod
  • MLOps - Docker, Kubernetes, GCP, AWS, Weights and Bias, Distributed Training, Tensorboard
  • Project Management and Tools - Git, JIRA, Confluence, Slack, Pycharm, IntelliJ, Jupyter Notebook

Research Works and Developement

A Style-Based Generator Architecture for GAN   Paper   Slides  

General Adversarial Networks(GAN) networks are one of the most used Supervised Learning. Given a training set, this technique learns to generate a new data with the same statistics. It contains two networks Generator, which generates new data and Discrimintaor, which evaluates them. Style GAN's are new proposed alternative method of GAN enhanced by style transfer literature

Understanding of VAE and GAN   Paper   Slides  

Generative adversarial network (GAN) and Variational autoencoder (VAE) are two commonly used deep generative models that can generate complicated synthetic images. This research includes different variations of Models like Vanilla GAN(DCGAN), Wasserstein GAN (WGAN), Vanilla VAE and VAE-GAN

Understanding of Tensorflow Tensorboard   Slides  

This research includes the understanding the concepts of TF Tensorboard and its applications

Understanding of RNN Networks   Slides  

This research includes the better understanding of different RNN Networks like RNN, LSTM, GRU with architecture, shapes and operations of each block inside these cells

Interests

I possess a deep passion for coding, mathematics, and logical thinking. Alongside my technical abilities, I have honed my teaching skills through training freshers in my previous company and teaching Python to friends. I actively participate in coding competitions on platforms such as Codechef and Leetcode, seeking to continuously enhance my skills.

In addition to my interest in technology, I am an avid cricket fan, closely following the ICC and BCCI, as well as other sporting events around the world. During my free time, I indulge in watching TV series, and thanks to the pandemic, I have discovered and enjoyed a multitude of new shows.

Awards & Certifications

Awards
  • Awarded 3rd Prize in Coding competition in the event "CODE DRIFT" conducted by Department of Computer Science and Engineering, ANITS.
  • Awarded a 2nd Prize in 36th Maths Olympiad conducted by Association of Mathematics Teachers.
Certifications
  • Participated in various Paper Presentations conducted by Computer Society of India(CSI).
  • Participated in "Shristi-Mobile Making work shop" conducted by Pragnan 2013 the International Techno - Management Festival of National Institute of Technology, Tiruchirappalli,India

Contact Me