Hello, my name is
Rajarshi Roy
And I'm a

About me

I'm Rajarshi, a

Hello! I’m Rajarshi Roy, a Computer Science and Engineering student at Kalyani Government Engineering College, where I maintain an 8.8 CGPA. My journey goes beyond academics—I've gained hands-on experience in AI/ML research and engineering, applied theory to real-world impact, and delivered results under high-pressure environments. I am passionate about developing innovative solutions and advancing technology through both research and practical projects.

  • Winner, Smart India Hackathon 2022 — led Team BRAINCELLS to deliver a 36-hour education solution adopted during COVID-19.
  • AI Engineer Intern at Qest — designed LLM-powered onboarding agents, developed session persistence, and enabled intelligent data retrieval using MongoDB.
  • AI Research Intern at Artificial Intelligence Institute of South Carolina — mentored by Prof. (Dr.) Amitava Das, contributed to groundbreaking NLP research and helped organize AAAI’25 Defactify-4.0 shared tasks.
  • Published and accepted research at top-tier venues (ICCV-CVAM, ACL Findings) on multimodal LLMs and kernel-based preference optimization.
  • Skilled in data analytics, machine learning, web development, MLOps, and cloud deployment (TensorFlow, PyTorch, Scikit-learn, FastAPI, Docker, AWS, Langchain).
  • Certified in Machine Learning and Google Data Analytics.

Explore my portfolio to discover the intersection of academic excellence and practical innovation. Let’s connect and shape the future of technology together!

DOWNLOAD RESUME

Publications and Ongoing Research

June 2025

ByDeWay (CVAM Workshop @ ICCV, Accepted)

  • Proposed a training-free framework using Layered-Depth-based Prompting (LDP) to enhance spatial reasoning in Multimodal LLMs.
  • Improved F1-score by up to 10% on hallucination-sensitive and reasoning tasks (POPE, GQA) across GPT-4o, Qwen2.5-VL, ViLT, and BLIP.
June 2025

DPO-Kernels (ACL Findings, Accepted)

  • Enhanced the DPO framework by integrating diverse kernel representations (polynomial, RBF, Mahalanobis, spectral) and embedding-based hybrid loss functions.
  • Proposed a data-driven kernel-divergence selection mechanism and evaluated it across 12 datasets, achieving state-of-the-art performance in factuality, safety, reasoning, and instruction following.
June 2025

DETONATE

  • Introduced DETONATE, a large-scale benchmark (~100K prompt-image pairs) for evaluating alignment across social axes (race, gender, disability), and proposed the Alignment Quality Index (AQI) to measure latent-space separability.
  • Extended Direct Preference Optimization (DPO) with kernel methods (RBF, polynomial, wavelet) and hybrid loss formulations, demonstrating robustness with divergences like Wasserstein and Rényi.

Work Experience

Apr 2025 – May 2025

Qest

AI Engineer Intern

  • Designed and developed a 3-stage LLM-powered onboarding agent that analyzes business data using Google Maps place_id, determines eligibility based on supported categories, and dynamically generates personalized platform components.
  • Implemented session persistence to allow seamless resumption of partially completed onboarding flows.
  • Contributed to the early development of a MongoDB-integrated agent to support intelligent, structured data retrieval for downstream LLM workflows.
May 2024 - Present

Artificial Intelligence Institute of South Carolina

AI Research Intern

  • Under mentorship of Prof. (Dr.) Amitava Das, applied theoretical understanding to practical solutions in NLP.
  • Contributed to a publication (under submission) on integrating diffusion methods with reinforcement learning for preference optimization.
  • Appointed as Defactify-4.0 Workshop Web Chair at AAAI'25, helping organize shared tasks on Codalab and supporting participants.
  • Provided strategic guidance and expert support to participants, enhancing their experience and ensuring the successful execution of shared tasks throughout the competition.

Projects

AI Research Assistant

Developed an AI Research Assistant app utilizing Gen AI and RAG model, powered by Google Gemini. Facilitates question answering from uploaded research papers via Streamlit. Data is cleaned, indexed using Llamma-Index with the Gemini Embedding Model.

Tech Stack : Python, LlamaIndex, Google Gemini, Streamlit, Pinecone, Docker, Github Actions, AWS ECR, AWS EC2.
Game Recommender App

A game recommendation system, powered by Google Gemini, offers seamless scalability and natural language query handling. Its user-friendly chatbot interface built with Streamlit integrates user feedback for ongoing enhancement. Langsmith facilitates logging and monitoring.

Tech Stack : Python Langchain Google Gemini Streamlit FAISS Docker Github Actions AWS ECR AWS EC2 Terraform

My Sweet Home

Engineered a housing price prediction system employing content-based filtering. Achieves 90% accuracy for sales and 70% for rentals. Offers 6 similar property recommendations with cosine-similarity percentages.



Tech Stack : Python, Flask, HTML and CSS, Scikit-learn, DVC, MLFlow, Docker, Github Actions, AWS ECR, AWS EC2
Check Out More Projects... Check Out My Certifications...

My skills

My creative skills & experiences.

Hello there! I'm currently diving into the world of AI and Machine Learning & excited to use my skills to solve everyday problems through cool projects

Know more
Languages
Python, SQL
Data Science
 Machine Learning, Deep learning, LLMs
Mathametics for ML
 Statistics & Probability
MLops Tools
DVC, MLflow, github actions
Databases
 Mysql, Mongodb, Postgresql, Pinecone
Gen AI Framework and Models
LlamaIndex, Langchain,
Google Gemini

Achievements and Blogs

Smart India Hackathon'22 winners

Our team "Braincells" pitched a solution for the problem statement "Suggest an innovative approach to serve the educational needs of the weaker section of the society during this challenging situation of COVID-19" provided by the Ministry of Housing and Urban Affairs . We developed an app having two versions-one through which online classes can be conducted during pandemic and the other one through which people can volunteer to educate and donate the weaker section of the society through offline events. We also gamified our app to make the volunteering-teaching-learning cycle interesting.

Navigating the Rental Maze: Crafting a Coding Adventure and Landing My First Kaggle Medal

Explored rent costs in Bangalore and Hyderabad, sparking a coding venture. Employed Scrapy and Selenium for web scraping, built a housing price prediction project with Krish Naik sir's guidance, and added a personalized recommendation feature. Check out the GitHub scraping project and Kaggle dataset. Secured my first Kaggle bronze—just the beginning!

Contact me

Get in Touch

Looking forward to hear from you !!
Let's connnect

Name
Rajarshi Roy
Kaggle Account
Email
royrajarshi0123@gmail.com
LinkedIn
GitHub