ML Engineer | NYU Courant, IIT Guwahati

Abhinav Gupta  • Peerlist

Explored domain adaptation and few-shot learning capabilities of semi-large PLMs like GPT-2 and CTRL, to improve performance on low resource downstream tasks without having to deal with large computation power needs.

Domain Adaptation - NLU Course Project

Built a chrome extension which could remove posts from facebook news feed (live) which are related to certain tags provided by the user. Extension uses NLP API to perform categorisation of posts.

Facebook News Feed Improver

Worked on human pose transfer problem using a GANs.

Generative Adversarial Network

There are a lot of software requirements specifications included in the functional requirements of the Hospital Management System, which contains various processes, namely Registration, Check out, Report Generation, and Database.

Hospital Management System

Built an app which could scrap out the text from an image using Microsoft Computer Vision API and store it or append it to any file. The text obtained can be translated to any other language using Microsoft Bing Translator.

Imagica (Microsoft Code-Fun-Do)

Made a Football playing Intelligent Robot, using  "Image Processing (Live Video Input) and Raspberry Pie". Task was to enable the bot to detect a specific colored ball in the environment/field, move towards it (2-D motion) without crossing half-court, align in correct position and score a goal in presence of opponent team bot.  Camera was placed overhead. Each section was colored differently for proper threshing.

Kick Ball Theory

Worked on creating a public benchmark suite for Emergency Departments (ED) using a large EHR dataset, focusing on three ED-based outcomes: hospitalization, critical outcomes, and 72-hour ED re-attendance.

Machine Learning in Healthcare

Developed a social media app to comprehensively understand the MERN (MongoDB, Express, React, Node) Stack.

MERN-Based Full Stack Project

Analyzed and implemented several deep learning methods to recognize drug-drug interactions (relation extraction) that appear in clinical texts like DrugBank and MEDLINE.

Relation Extraction Survey - NLP

Secured first place on the leaderboard in a team-based competition focused on finding the best self-supervised learning algorithm for object detection. Our best-performing model was Barlow Twins with a Deformable DETR head.

Self-Supervised Learning for Object Detection

This interpreter is based on Python Object Oriented Programming and uses auto-linking of objects to their functions (saying in lame terms). All the "__init__" methods and "build" methods are used to parse the blocks of code sequentially. The "eval" method in the above interpreter works similar to python's "eval" method and evaluates the parsed tree/structure of the code.