Jan 2021 – Jul 2021 · 6 mos · Bengaluru, Karnataka, India

• Conducted a detailed and comprehensive survey on the SOTA for improving the throughput on the inference task of transformer models.
• Proposed and tested 2 novel architectural innovations for improving the latency of transformer-based language models.
• Conducted extensive experiments using various modelling approaches on an internal dataset to benchmark their speed-accuracy trade-off.

Jul 2020 – Dec 2020 · 5 mos · Bengaluru, Karnataka, India

• Team: Wireless Tools and Automation
• Worked on the mainline development of an Internal Trace and Log Analysis tool.
• Tech-stack: Swift, Cocoa, Objective-C, XCode, SwiftUI, AppKit

May 2020 – Jun 2020 · 1 mo · Gurgaon, Haryana

• 1. Visual Object Segmentation
• Research oriented problem focusing on finding scalable techniques for instance segmentation. Implemented YOLACT, SOLOv2, BlendMask networks using PyTorch, and analyzed the performance results.
• Managed to obtain a speedup of about 20x compared to the currently in production Mask-RCNN model, making the system more scalable for real-time processing of images.
• 2. Smart Text Tagger and Visualizer
• Developed a web platform for tagging large datasets (in millions of records) efficiently using unsupervised techniques like Clustering, Rule Engine.
• Used HTML, CSS, JS for frontend, Flask, Celery for the backend. The database was hosted on MongoDB.
• Added support for multiple user collaboration by designing protocols to ensure consistency. User and group permissions are used for authorization.

May 2018 – Jul 2018 · 2 mos · Bengaluru Area, India

• Trained a neural network on a large-scale fashion dataset (DeepFashion).
• Extracted features from the neural network and used them to calculate similarity scores between images.
• Built an image search engine using the Euclidean similarity scores and hosted it on Google Cloud using the Flask Framework
• Built an object detection module using Tensorflow to detect and localize fashion products in a video and show relevant advertisements in a banner overlay using OpenCV.
• Developed an android application for a smart TV that allowed the viewers to search for visually similar products in real-time by making PULL requests on their TV screens for items of interest. The application detects items on the screenshot of the requested video frame and searches for visually similar products.