cv

Graduate Researcher @ Microsoft

Remote, US

• Researched zero-cost performance evaluation for neural architecture search (NAS) on automatic speech recognition.
• Faster performance evaluation speeds up the architecture search by multiple folds. Also experimenting on how learnings transfer across domains (e.g. how well zero-cost proxies learned from vision task transfers to speech recognition).
• Technologies used
  • Python, Slurm, Matplotlib, Singularity, Tensorflow, Pytorch, Numpy, AutoML

Data Science co-op @ Ericsson

Santa Clara, CA, USA

• Performed research on active learning for labelling time series data. Worked on novel proposal utilizing attention to build context of the unlabeled data before sampling from them. Also explored reinforcement learning to choose best sampling method given current set of labeled samples.
• Collaborated on evaluation of graph neural networks for identification of potential cells to put to sleep, thus leading to reduction in overall energy consumption. Helped the team explore tree-based alternatives with a fraction of hardware requirements.
• Technologies used
  • Python, Docker, Scikit-learn, Matplotlib, Numpy, Pandas, Gitlab, Jupyter

Data Scientist @ Flam

Mumbai, India

• Researched and led development of algorithms to generate context-sensitive advertisements for users of the social platform.
• Context creation was aided by user features and possible interests based on their social interactions. These context-enhanced algorithms, along with relevant target advertisers, helped increase clickthrough rate by over 3x. Parallelly worked on system design and deployment for the same.
• Technologies used
  • Python, MongoDB, Flask, Tableau, RabbitMQ, Matplotlib, Amazon Web Services, TensorFlow, Numpy, Pandas

Researcher @ Indian Institute of Technology Bombay

Mumbai, India

• Worked on solving computer vision problems using deep learning. These problems included statements like automated facial re-enactment using generative adversarial network (prototype available on Github), text spotting and automatic numberplate detection.
• Simultaneously led development of “Lokacart”, a NFPO e-commerce platform aimed at facilitating direct business between farmers and customers. Worked on (and later led) development of API platform, product recommendations, payments, and data management portal.
• Technologies used
  • Python, Java, Matplotlib, Google Cloud Platform, Node.js, Spring Boot, MySQL, Scikit-learn, Numpy, Pandas, Tensorflow

MS in Computer Science

University of Massachusetts, Amherst

• 3.8 CGPA
• Concentration in Data Science
• Courses
  • 3D Vision Computing
  • Reinforcement Learning
  • Neural Networks
  • Statistics
  • Advanced NLP
  • Machine Learning
  • Computer Vision
  • Secure Distributed Systems

BE in Computer Engineering

University of Mumbai

• 8.64 GPA
• Courses
  • Software Engineering
  • Object-Oriented Programming
  • Analysis of Algorithms
  • Computer Networks
  • Distributed Databases
  • Operating Systems
  • Data Warehouse & Management
  • Image Processing
  • Big Data Analytics

gRPC based trading application

• Created a fault tolerant and distributed gRPC based server to facilitate trading between multiple clients and servers, simulating a real-time distributed system.
• Also experimented with dynamic Threadpool implementation built from ground up to compare performance with existing mutithreading libraries.

Semantic segmentation of 3D point cloud using sparse representation

• Performed semantic segmentation on S3DIS dataset using a modified UNet3D architecture with residual and attention layers.
• Sparse point cloud representation helped reduce memory footprint. Obtained an average mIoU of 0.55 compared to 0.41 of SegCloud.

Evaluating detection-free object tracking

• Compared the performances of E2E detection-based and detection-free tracking in terms of accuracy and runtime.
• Experimented with k frame skips between detections and observed over 2.7x reduction in runtime (k=5), showing viability of tracking algorithms on compute constrained devices.

Self-supervised learning of image representation by solving Jigsaw puzzles

• Implemented a paper on learning image representations by using solving a pretext task (Jigsaw puzzle).
• A Siamese network with shared parameters was trained on all (9) pieces with a dense layer combining the outputs.
• Changed the base architecture to Resnet50 resulting in mAP of 40.2 compared to reported mAP of 41.8 despite minimal compute and training.

Facial re-enactment using Generative Adversarial Networks (GANs)

• Researched replication of facial structures from a video via facial reenactment for automatic dubbing.
• Performed alignment of face with an approximate 3D model to map lip movement from a source frame onto a target.
• Generated patch for lip sync by using conditional GANs, allowing for a single model irrespective of facial structure.

Automatic number-plate recognition

• Contrived vehicle detection in each image to help increase accuracy of number plate recognition, which was made accessible using REST APIs.
• Implemented detection of vehicle with You Only Look Once (YOLO) object detection, followed by text orientation and scanning with FOTS to improve region-of-interest detection by up to 6%.

• Grading assistant for 3D visual computing course (COMPSCI 574/674) at UMass, Amherst

• Ranked in Top 2% of Morgan Stanley Hackathon across India
• Represented the Computer Society of India student chapter as Sr. Technical Officer

• Qualified for Indian National Math Olympiad