Interests
I am interested in areas of systems, cloud and serverless computing, machine learning, and the intersection of these. Currently, I am doing Research and Development on performance characterization of software-assisted GPU scheduling and scheduling in FaaS platforms for GPU functions with variants.
I am also working in areas of speech processing, improving the preprocessing stages that heavily affect the accuracy and model design decisions of ML-based speech applications.
Some cool projects
Ongoing
Performance characterization of GPU multiplexing using MPS
I am working with Prof. Puru and Prof. Umesh Bellur on empirical analysis of NVIDIAs MPS enabled GPU multiplexing and benchmarking interference effects of MPS when co-locating multiple processes on GPU.
Speech-text alignment
As part of my BTP with Prof. Preeti Rao, I am working towards evaluating the methods for preprocessing broadcast news audio of low-resource languages and improving them for use in ML-based speech-to-text translation models. We implement approximate string-matching algorithms to find string similarity.
Past
SeDriCa
Sedrica is a student-led technical team developing India's first self-driving car with level 5 autonomy. I worked in the computer vision subsystem for about a year and developed a multi-task learning model for jointly performing object detection and road segmentation. We significantly reduced computation costs on limited GPU memory using this common backbone and multi-head design.
Automatic Speech Recognition
As part of my course work CS753 - Automatic Speech Recognition under Hacker role, we implemented the paper Right2Talk on main speaker identification and localization using audio-visual transformers. I also performed transfer learning on the Coqui English STT model for building ASR systems for Indian languages on extremely small datasets.
Low Light Image Enhancement
We all have faced that situation when we wished we could somehow make the night images brighter or enhance images captured under low light conditions while ensuring they look natural. This motivated me to implement and compare different low-light image enhancement techniques in my Image Processing course project—details in the report
Bosch's Traffic Sign Recognition Challenge
Won bronze out of 23 teams that participated in Inter IIT Tech Meet 2021. The problem statement involved creating a no-code UI for all the stages of the ML pipeline while also providing intelligent solutions to the user to explain failures of the system. I particularly worked on the backend, training different models on the challenge dataset and implementing different visualization techniques for model explainability.
Image Super Resolution
I implemented a Generative Adversarial Network (GANs) to realistically super-resolve the low-resolution images on the DIV2K dataset. The low-resolution images were synthetically generated using bicubic downsampling.
The Tracking and Navigation Challenge
Under Autumn of Automation, UMIC IIT Bombay, I programmed a bot with ROS to solve a perfect maze while avoiding obstacles in the path. The problem statement also required the bot to recognize letters on the wall and permute a code to crack the safe.
Low cost POF link
In this project, we built and tested Polymer Optical Fibre offering a lightweight, low-cost and broad bandwidth as an alternative to glass optical fibers for digital transmission across a length of about 10m and achieving data rates of 10Mbp. We designed a PRBS transmitter and trans-impedance amplifier-based receiver on PCB.