MedLam - An Intelligent Medical Assistant

Open sourced a medical Q&A LLM at Hugging Face, driving 5000+ downloads by finetuning Meta’s Llama using LoRA PEFT approach.

Improved ROUGE and BERT score by 20% on medical data by designing and building a data-processing pipeline for 100K records.

Tech Stack:Python, PyTorch, Spark, Machine Learning, Large Language Model, NLP, Deep Learning, Data Analysis

Distributed Key-Value Store Using Raft Consensus Algorithm

Ensured strong consistency, linearizability, and state recovery by developing a fault tolerant, replicated key value store using Raft.

Achieved leader failover to 4sec by sending 5 − 6 heartbeats/sec for leader election and implementing persistent log replication.

Tech Stack:Golang, Linux

Real-time Group Chat Server

Enabled real-time messaging via websockets with 99% uptime, handling up to 100 live group members, making it a multi-client server.

Dropped message latency by 30% by assigning each user to a new thread, ensuring concurrent communication between server and client.

Tech Stack:Java, WebSockets, Muti-threading

COVID-19 Information Hub

Developed this project during the COVID-19 pandemic to raise awareness and support for the community by keeping them informed with the latest news and data regarding COVID-19.

This website displays the latest news from popular broadcasters and critical tweets worldwide. It performs analysis on massive datasets that are updated daily to show relevant information in charts and tables for users to visualize the change in cases, deaths, and vaccines administered worldwide.

Tech Stack: C#, .NET, JavaScript, HTML5, CSS3, Entity Framework, LINQ, MS-SQL

Trajectory Generation BERT

This scalable trajectory imputation system was developed to address the challenge of sparse and incomplete trajectory datasets. Numerous applications, like pathfinding, traffic monitoring, location-based services, map inference, and urban planning, rely heavily on accurate trajectory data generated from GPS devices. However, due to limitations like bandwidth, storage constraints, etc, trajectory coordinates are sparse, leading to reduced accuracy.

This framework inserts realistic coordinates in a trajectory to enhance accuracy by mapping the trajectory imputation problem to the natural language processing(NLP) technique of finding the missing word by leveraging the widely used BERT LLM. It adapts to the unique characteristics of trajectories and doesn't require knowledge of the underlying road network or dense trajectory data.

Tech Stack: Python, NumPy, Pandas, Jupiter Notebook, Spark, BERT

70 Years of Music Visualization

This project seeks to explore the rich data of music, aiming to uncover hidden patterns and trends, tell stories about the past, and help us understand how music has changed. Through effective visualization techniques, it offers a unique perspective on the world of music.

Tech Stack: Python, JavaScript, D3.js, HTML5, CSS3

DNS Resolver

This is a DNS Resolver that caches the previously requested domain for a certain period of time for faster processing of the requests made for the same domain name. It uses Java's multi-threading feature to process mutiple requests parallelly.

Tech Stack: Java

Unix Shell

In this project, I crafted a simple Unix shell in C++. This custom shell allows users to run commands, incorporating features like I/O redirection and piping, akin to common Unix shells such as bash and zsh.

Tech Stack: C++

Face Recognition

This is a web application that locates a face in the image. A user must register and log in to the application to keep track of the total number of requests made to the application. It uses machine learning APIs to locate the face in the image and display it in the UI.

One of the real-life uses of this feature is in human verification through an image-based captcha.

Tech Stack: React, Node.js, Express.js, PostgreSQL

MSD Script Interpreter

This is a scripting language interpreter developed using C++ that interprets the MSD Script programming language. MSD Script is a case-sensitive programming language that can perform complex operations such as numerical calculations, variable declaration, initialization, and function creation and invocation.

Tech Stack: C++

Machine Learning Library

This project is a lightweight machine-learning library that implements some of the well-known classical machine-learning models, such as Decision Trees and ensemble methods like AdaBoost, Bagging, and Random Forest. Other models include Perceptron, SVM, Logistic Regression, and Neural Networks.