Work Experience

SAP Labs

Software Developer

Aug '18 - Sep '21

I served as a full-stack developer at SAP SuccessFactors, a company specializing in cloud-based Human Experience Management solutions for businesses. My primary role was to develop new features and enhance user experience by providing customers with an innovative and highly customizable Recruiting platform. In addition to implementing new features, my role also involved ensuring the performance of the product and reducing defects in the code. In that regard, some of my accomplishments include leading the integration of Containers and Kubernetes to transform the Agency product into a microservices-based architecture and reducing customer defects for Agency by 50% using Test-Driven Development.

Data Science Intern

Jan '18 - Jul '18

I was responsible for enhancing user experience for a key client by integrating Machine Learning into their products. I developed a Speech-to-Text API using probabilistic models, leveraging Hidden Markov Models and LSTMs to accurately recognize diverse accents and pronunciations. In a subsequent project, I implemented a Random Forest classifier model to categorize mobile app users based on extracted features. This enabled our team to pinpoint areas for improvement, ultimately driving increased app engagement.

Avagmah

Software Development Intern

Jun '16 - Jul '16

In the summer of 2016, I interned at an Ed-Tech startup, marking my initial foray into the industry. I was responsible for crafting complex SQL join queries to generate insightful reports for the customer analytics dashboard. Additionally, I redesigned the company's homepage utilizing HTML, CSS, and JavaScript. I also developed the end-to-end functionality for the 'Recent New Videos' Moodle block using PHP.

Projects

AffordanceNet

Instance Segmentation to detect the affordances of different parts of objects in an image. The code uses the Segformer MIT-B0 architecture as the encoder to learn the feature embeddings of the image. The embeddings of the SegFormer encoder is passed to 2 CNN decoders. One decoder detects the bounding boxes of the object and the other performs pixel‑wise affordance detection of the objects detected trained with a joint loss function.

Reinforcement Learning: Gridworld using PPO and Behavioral Cloning

Using the concepts of Reinforcement Learning to solve the Gridworld problem. The code employs Proximal Policy Optimization (PPO) coupled with Behavioral Cloning using datasets containing the initial state and final state of the Gridworld environment to learn the policy. The reward function is tweaked to create a balance between an aggressive and a successful policy. Behavioral Cloning helps accelerate the learning by providing guidance to the model especially in the initial phases of training. Using this combination of On-policy methods along with Behavioral Cloning allows for a more stable optimization for higher order MDPs and ensures that the policies improve steadily.

Semi-Supervised Image Classification

An implementation of Semi-supervised Image Classification on the CIFAR10 and CIFAR100 datasets to show the scalability of deep neural networks in the absence of large datasets of labeled data. This project implements 3 common techniques for SSL:

  • Pseudo-label - Dong-Hyun Lee et al.
  • Virtual adversarial training - Takeru Miyato et al.
  • A modification of Fixmatch - Kihyuk Sohn et al.

Speech To Text API

API to convert speech to text. Built using Flask Framework taking input as Base64 encoded sound recordings from the client-side and outputs the detected speech as a string back to the client-side.