Bachelor's Thesis
Thresholding Binarized Neural Networks to Improve Accuracy in Large Language Models Training
Undergraduate Thesis – 2024 | NIT Trichy, India
Under the supervision of Prof. N Ramasubramanian, my team and I worked on improving accuracy during a large- language model training while reducing the model size. We developed a thresholding mechanism that compressed the model while enhancing its overall performance.
Work Experience
Advanced Machine Intelligence Research (AMIR) Lab
Research Assistant
February 2025 – Present | Bangladesh
Investigating LLM reasoning capabilities to improve performance on complex tasks, with a focus on reducing hallucinations and achieving human-like decision-making.
Advised by Dr. Mohammad Firoz Mridha
Indian Institute of Technology Bombay
Research Intern
May 2023 – July 2023 | MeDAL Lab | India
Worked on enhancing a Self-Supervised Learning Framework (BYOL) for better representation learning. Created a user friendly system for DICOM image anonymization.
Advised by Prof. Amit Sethi & Gouranga Bala (PhD)
National Institute of Technology Tiruchirappalli
Undergraduate Reasearch Assistant
October 2022 – February 2023 | Industrial Automation Lab | India
Developed a Flutter app to automate the operations and billing for a power cluster in Assam, India, which serves around 200 homes. This project was funded under SUSTENCE , a Govt. of India initiative for carbon-neutral energy communities.
Advised By Prof. M. Brindha & Prof. G. Saravana Ilango
Selected Projects
Retrieval Augmented Generation (RAG) Powered Cold Email Generator
April 2025 | Code / Instructions
Developed a Python and Langchain-based RAG system that generates personalized emails matching graduate applicants with professors based on research alignment. This pipeline leverages Llama-4-Maverick chat model (Groq AI), Jina AI Embeddings model and Cohere's Reranker to generate highly contextual cold emails for improved response rates.
Tech Used: Python, LangChain, FAISS (Vector DB)
NoSmokeZone: AI on the Hunt for Smokers
December 2023 – January 2024 | Code / Instructions
Trained advanced CNN architectures (VGG16, ResNet-50, EfficientNetV2) and Vision Transformer for smoker detection to enhance public health surveillance. Conducted a comparative analysis to evaluate model effectiveness and identify areas for improvement.
Tech Used: Python, TensorFlow, Keras, OpenCV
C Compiler for while loop with nested while loop and/or if-else constructs.
March 2023 – April 2023 | Code / Instructions
Built and automated a simulation of the front-end and back-end phases of a C compiler, focusing on If-Else, While, and nested If-Else-While constructs. It covers six phases of a compiler. This project was part of an academic course Compiler Design (CSPC62).
Tech Used: C, Yacc, Python, Shell
9 Box Puzzle – Android App
December 2022 | Code / Instructions
Developed a 9 box puzzle game in Java with a 3x3 grid, featuring an AI-driven solution using the A* algorithm. The game includes a home menu, how-to-play guide, optimal solution visualization, and a performance tracker based on user statistics and puzzle difficulty.
Tech Used: Java, Android Studio, A* Algorithm
Malicious Website Detection using Machine Learning
September 2022 – November 2022 | Code / Instructions
Trained a machine learning model with 94% accuracy to detect malicious websites. Built a Chrome extension using JavaScript to extract webpage features and classify sites, incorporating user feedback for continual learning. The model retrained daily, improving accuracy over time.
Tech Used: Python, Flask, HTML/CSS, JavaScript, Chrome Extension Tools
Cube Runner
August 2022 | MacOS Build | Demo / Video
This is the first game I created using Unity. It's a low budget Temple Run but with a cube. It has 3 levels. With each level the the speed of the cube increases which makes it harder in the last round to finish up. Inspired from - Brackeys <3
Tech Used: Unity, C#
For the complete project list, check out my GitHub profile. Apologies if you're unable to access any projects or encounter a broken link, the repository may have been made private.