Hateful Memes Challenge

Python, PyTorch, Hugging Face

Created a model to compete in the Hateful Memes challenge by Meta. The model utilizes Facebook’s MMF (Multi-Model Framework) and the MultiModal Bitransfomer HuggingFace Model, which integrates BERT and CLIP models. The model uses the CLIP part of the model for image encoder and BERT part of the model for text encoder. There was an extensive tiling function applied for the image processing that enabled more context for the CLIP model. The text was augmented using the Easy Augmentation Method. This model with relatively low short amount of training managed to make the same performance as the top 50 submissions in the competition.

Tectonic (WIP)


Contributed to an open source project called tectonic, which is a modernized, complete, self-contained TeX/LaTeX engine, powered by XeTeX and TeXLive. I added the ability for users to specify a target output pdf to build, amongst a series of outputs.

BuzzOS Rust Standard Library


Worked on making a minimal implementation of the Rust Standard Library for BuzzOS, which is an OS developed by the TAs of the Operating Systems class at GT. This special version of the Standard Library was implemnented without a libc, and using wrapper based interface around raw syscalls. I included a String, Vector, Box and Cell Implementation along with the multiple traits implemented for each of these types. Implemented a user space, and allowed the kernel to gracefully handoff to the userspace after setup.

Mandelbrot Madness (WIP)


Mandlebrot Madness is a 3D Fractal Based Platformer that is written in C++ using the SFML library. Writing a 3D camera with mouse and keyboard controls for the user.

Brush (WIP)


Brush is a minimally POSIX compliant shell that is written in C. Written to implement how much ever of the POSIX compliancy I can. Added basic shell that can process any command in the PATH with proper error reporting. Working on implementing a POSIX interpreter enabling the building of command-line scripts.


Python, ReactJS, Express.js

Speech2Code is a turing-complete website that helps translate a spoken english like meta language to Java, JavaScript and Python, using a language agnostic class-based abstraction system that enables easy addtion of language support. The speech transcription is done using the AssemblyAI API. I worked on making the Programming Language Middle-End and worked on making CSS changes on the front-end website.

DetectCNN (WIP)

OCaml, Owl, Eigen

DetectCNN is my implementation of the Mask Region Based Convolutional Neural Network Paper, written in OCaml using Owl and Eigen. Worked on sorting LLVM issues in Owl on Darwin M1 by contributing to the Github issue. Read and summarized findings of the R-CNN, Fast R-CNN, Faster R-CNN Papers. Studied Instance and Semantic Segmentation approaches, Region Proposal Techniques and Feature Extraction Methods.


Python, Tensorflow, Keras

Enhance is my implementation of the Super Resolution Generative Adverserial Network Paper using tensorflow, which uses a ResNet with a Perceptual Loss function based on discriminator probability and the euclidean distance between the feature representations of a reconstructed image and reference image, to achieve a high level of super resolution. Ran and trained the model on the DIV2K dataset achieving perceptible improvements in image quality on the DIV2K dataset with downscale factor 4.


Rust, Yew Framework

DeusCL is a memory-safe Common LISP implementation with a REPL built in Rust . Built a handwritten recursive descent parser and a visitor pattern dependent evaluator. Added a basic type inference, type checking and error handling system. Added dynamic evaluation of functional environment and parameters. Wrote a simple garbage collector and environment cleaner. Built a frontend in Yew Framework in Rust to enable users to use a web based simulation of REPL.


Unity, C#

Working on Paper Planes, a 2D sidescroller platformer created with Unity as a part of the VGDev Club on campus. I worked on a relaunch mechanism using custom aero dynamics, a particle system for paper plane trail, and one for collisions with objects. I also worked on making a Quest API to show the current quests going on.



As a part of a dependently-typed (Programming Language Club) on campus, I worked on a python implementation of the common teaching language Lox. I worked on building and degubbing a Recursive Descent Parser, Scanner and a made a script to generate a python class system for different atoms and a visitor pattern for the interpreter class. I also worked on the AST and a language environment to store function and variable declration. I lectured at the event about AST evaluation, dynamic environment creation for functional and recrusive calls and implementating the vistitor patters for interpreters.


ReactJS, Express.js, Flask, Firebase, Solidity, Ganache, IPFS, Web3.js, NCR API

Publixmon is a full-stack mobile app integrating modern day NFT and Style Tranfer technologies to create a unique interactive retail experience. With every checkout, the app has a chance to give the user a NFT using a style tranfer machine learning model based on two attributes a famous painting by a renaissance painting and an item in your purchase history. Users walk into a store and scan a QR code to see other active users who are in the store and they can socialize, see each other’s collections of virtual toys, and trade them to grow their collection. A user’s collection is auto-scored based on size and variety of collectibles, resulting in higher engagement for the business. I worked on making the backend blockchain that generates the NFTs that uses the style tranfer images using Solidity and a Ganache Blockchain to test it on. I worked on creating the backend for the project that uses the NCR API to serve purchase history and CDN Network to store the style transfered images.

Mental Health Chat

ReactJS, Express.js, MongoDB, Firebase, Socket.IO

A timed chat for people to share and talk about issues they’re having in relation with their mental health and help them overcome it. A message in the Redis database to expire in about 6 hours in order to encourage people to come out and discourage users from harrasing one another. Created a Dockerized App with a Redis Instance during development. Set-up a Log-In page using Firebase that is capable of using Google OAuth, the users can then set their username, add a profile picture and set the theme (light/dark) for the app, before getting started chatting on the server.


ReactJS, Express.js, MongoDB, Firebase, Twitter API, Gensim, NLTK, Node.js

A full-stack fake disaster tweet analysis tool that uses TwitterAPI to extract the text from the tweet link that the user pastes in input box and data processing tools like **lemmatization, tokenization, N-Grams, Stop Words **, removal of punctuation, and a TD-IDF based Logistic Regression NLP model with a 81% accuracy. Used MongoDB, to store user hashed user credentials, created a token authentication system and used localstorage to automatically sign-in and verify the user.