About Me

I am a data science enthusiast and an active data science mentor with Masters degree in Computer Science from University of Utah. I was a PhD candidate for a brief period of Spring 2017 to Summer 2019.
My work included deploying solutions for time-series and forecasting problems, data analytics and performance optimization for software verification tools. I like to read books, research papers and articles on Machine Learning. I also participate in Kaggle discusions whenever possible.

Experience

(Senior) Data Scientist, Arivo Acceptance LLC

• 35% increase in market capture rate by collaborating on updating the existing risk model to evaluate risk of default associated with new auto loan applications.
• 65% reduction in cost to verify income from credit bureau by building and deploying an income predictor for sub-prime customers for auto loan applications.
• Reduced customer deliquency rate by designing a linear programming optimization algorithm to help with portfolio management at collections department.

Data Science Mentor (MIT-IDSS), Great Learning

• Mentored over 300 students in collaboration with MIT for Applied Data Science Program (ADSP) and Data Science and Machine Learning (DSML) program.
• Instructed 60+ cohorts with an average feedback rating of 4.7/5.
• Topics include Hypothesis Testing, Regression, Classification, Deep Learning, Recommender Systems, Time Series Forecasting, Graph Neural Networks, AutoML tools.

Data Scientist, Aakash 88 LLC

• Increased annual profits by 13% by deploying time series models to forecast hourly load capacity and energy price for 150 wind farms in Texas to perform energy trading.
• Reduced manual analysis time by 35% through automation.
• Increased the volume of daily trades by monitoring and updating features in real time.

Machine Learning Researcher, University of Utah (SOARlab)

• Worked with SoarLab group under Prof. Zvonimir Rakamaric
• Given a program and set of properties, the objective was to use SMACK to provide satisfiability (Memory overflow, Concurrentcy, Memory safety, etc) and provide evidence for the satisfiability condition under 900s.
• I implemented an MLOps solution to check the satisfiability of the properties and implemented a subspace clustering and local search optimization to speed-up SMACK to reduce TIMEOUT errors. ALthough the focus was mainly to reduce the TIMEOUT errors, the implementation helped reduce Type II errors across all SV-COMP categories.

Graduate Research Assistant, University of Utah

• Prof. Suresh Venkatasubramanian (Summer 2014, Summer 2015)
• Prof. Vivek Srikumar (Summer 2016, Fall 2016)

Graduate Teaching Assistant, University of Utah

• Courses included Advanced Algorithms (Fall 2014, 2015), Data Mining (Spring 2015) and Discrete Mathematics (Spring 2016, Spring 2017).

Projects

• NLP Projects (Text or Audio datasets), (Code files)

  • SARAH: A locally hosted personal voice assistant with trigger word detection using an LSTM with attention model
  • Convert human readable dates into machine readable dates by using seq2seq model
  • Generate new jazz music by training a LSTM with attention model
  • Built character level language model to generate new names
  • Add appropriate emoji to text

• CNN Projects (Image based datasets), (Code files)

  • Generate new art using neural style transfer learning
  • Autonomous driving car detection using YOLO model
  • Facial expression recognition and verification model
  • Google Landmark recognition, Kaggle Competition
  • Self driving car speed prediction, comma.ai speed challenge

• Google Landmark Recognition 2019, Kaggle Competition (Read more) (Code)

  • Implemented a computer vision pipeline to detect landmarks in the training dataset of 4 million images using Python and Keras. Performed data cleaning to speed up the runtime 6x. Used transfer learning method with VGG16 for classification of the detected landmarks.
  • Training set: 4.2M images, Test set: 120k images, Classification categories: 203k

• Self driving car speed prediction, comma.ai speed challenge (Read more) (Code)

  • Extracted frames from video files and implemented a computer vision pipeline to predict speed of the car using Keras and OpenCV image processing capabilities.
  • Train set: 21k images, Test set: 10k images, Mean Squared Error = 3.2

• Automated prediction and analysis of SV-COMP benchmarks, SOARlab

  • Applied ML tools to analyze and categorize input C programs into  SV-COMP benchmark categories using Python, sklearn and Pandas  capabilities. Applied PCA, t-SNE to speed up the algorithm by 3x. 
  • Performed data labeling, data cleaning, data analysis and  hyperparameter tuning to optimize the accuracy and loss in the training  procedure. The developed model achieved 98% accuracy.
  • Training set: 6,000 benchmarks (approx. 5k loc each), Test set: 1,000 benchmarks, Classification accuracy: 98.02%

• CNN models (code)

  • Implemented code to classify images using different CNN architectures (GoogLeNet. ResNet, InceptionV3, DenseNet121, etc) leveraging the pre-trained data from ImageNet dataset.

• Pair trading and Machine Learning

  • Pair trading is basic form of Statistical Arbitrage. The concept behind pair trading is the assumption that two co-integrated stocks would not drift too far away from each other.
  • The documentation can be found here.
  • The code for this project is unavailable because I am planning to test this concept on live data-set.

• Simple Python Projects

  • English Dictionary
  • Website Blocker
  • WebMaps

Certifications

• Coursera

  • deeplearning.ai Deep Learning Specialization (5 courses) Link
  • Mathematics for Machine Learning Specialization (3 courses) Link
  • Google Data Analytics Specialization (8 courses) Link
  • Google Cloud Platform Big Data and Machine Learning Fundamentals Link
  • Machine Learning Link

• Udemy

  • Deep Learning A-Z: Hands-on Artificial Neural Networks
  • Machine Learning A-Z: Hands-on Python and R in Data Science
  • Data Science: Natural Language Processing
  • The complete SQL Bootcamp
  • Machine Learning, Data Science and Deep Learning using Python

Calendar

Blogs