Tanmay Ambadkar

I am a graduate student at PennState (PhD in CSE), advised by Dr. Abhinav Verma. I graduated from IIIT Vadodara with a B.Tech in Computer Science and Engineering. My works lies in the intersection of reinforcement learning and programming languages for safety and reliability of RL algorithms.

The Pennsylvania State University, University Park, PA, USA

• GPA: 3.77

The Pennsylvania State University, University Park, PA, USA

• GPA: 3.9
• Coursework: Neurosymbolic Learning, Vision and Language, Reinforcement Learning, Robust and Secure Deep Learning

Indian Institute of Information Technology, Vadodara, Gujarat, India

• GPA: 9.4
• Coursework: Topics in Deep Learning, Information Retrieval, Machine Learning, Artificial Intelligence, Data Analytics and Visualization,
• Clubs and Societies: President of Technical Committee IIITV, ML Lead of IIITV Coding Club and Developers Student Club IIITV, Pianist @ Encore - Music Club of IIIT Vadodara.

Dept. of Architectural Engineering, The Pennsylvania State University

• I am working on integrating reinforcement learning for building energy optimization.
• I worked on integrating dymola simulation tools with an open-ai gym interface, defining a robust multi-objective reward system and training multiple RL agents.

Dept. of Industrial and Manufacturing Engineering, The Pennsylvania State University

• I am working on predicting cases of Autism Spectrum Disorder in children within the age of 24 months, using Electronic Health Records (EHR) data, by using time-series models.
• Using PySpark, I preprocessed a dataset over 500GB to gather data for over 600000 patients. This data was used for cross-validation for multiple models to determine the best fit and most important features contributing to the final binary decision

Siemens Technology and Services

• I was responsible for detecting anomalies in data using AutoEnCoder models and using explainable AI (SHAP) to identify which features contribute to the anomalies. I developed a plug-and-play library for this having multiple layers of abstraction and obfuscation to protect intellectual property.
• I experimented with deep learning models and modifying workflows to create a library which could be used with any time-series data from multiple sources (csv, sql, influxdb). Demonstrated Proof-of-concpt using DataBricks on Siemens Buildings data to an internal team.
• I performed exploratory data analysis on coffee roaster data from Starbucks to identify burner cuts and its relation with other variables. I trained time-series models to predict burner cuts and developed a real-time prediction library to demonstrate its efficacy.

Siemens Technology and Services

• I worked on the Industrial Predictive Analytics Engine (IPAE), where I integrated workflow creation, management and progress monitoring using celery, redis and flask, which aided to running multiple workflows in parallel and reducing overall execution time of the pipeline by 30%.
• I was responsible for maintaining and upgrading a library for realizing model training workflows for time-series datasets. This was used in the Dubai Expo 2021.

The Pennsylvania State University

• I am working as a Teaching Assistant for the course CMPEN 270: Digital Design: Theory and Practice. I am responsible for conducting lab sessions, office hours, and grading exams.
• I was a Teaching Assistant for CMPSC221 Object Oriented Design & Web Programming. Conducted office hours and recitation sessions to strengthen OOP concepts taught in class, and graded projects for the course.

Shared Critic PPO for building energy and comfort optimization

• This project was part of EE597 - Reinforcement Learning. We developed a shared-critic PPO algorithm that optimises energy and comfort for a building using multiple agents. The simulator (sinergym) used for EPlus and bctvb to provide a gym-like environment for this task.
• Code | PPT | Report

Robust and Secure Deep Learning

• This repository has a set of 4 projects that were a part of CSE597 - Robust and Secure Deep learning. The projects involve designing test-time evasion attacks on the MNIST dataset and identifying them using a detector, and poisoning the dataset prior to training with an imperceptible pattern and then estimating that pattern using gradient descent
• Code

Deep learning for Strong Gravitational Lensing Images

• This project involves solving several tasks related to dark matter. I used convolutional neural networks and vision transformers for tasks like classification, denoising and superresolution.
• Code

Lead Scoring

• This project involved scoring leads for a marketing application. I performed exploratory data analysis and created a logistic regression model and a lead rating system that could be used by the marketing team to identify potential customers.
• Code

Certificate Generator

• I developed a certificate generator and repository Website for events of IIIT Vadodara. This sped up the process of delivering certificates to winners and participants and gave them an e-repository to verify the authenticity of certificates.
• I Coded the backend to manage mass-mailing of certificates and to generate a unique Code which could be used to identify the certificate.
• Code | Website

Gymkhana Website

• I developed the student gymkhana Website for IIIT Vadodara, where clubs and committees could showcase their activities and events.
• I developed a role management system for clubs and committees, and built a calender system to show upcoming activities.
• Code | Website