Research & Projects

The research focuses on model stealing attacks, particularly extracting parts of black-box production Large Language Models (LLMs), such as OpenAI's GPT-3.5 and Google's PaLM-2, through API queries. This work addresses the growing concern of protecting proprietary models while demonstrating the vulnerability of production LLMs. The attack successfully recovers critical architectural details like the embedding projection layer and hidden dimensions, showcasing the potential to extract sensitive model parameters efficiently and cost-effectively. 

Objectives

Investigate Model Vulnerability:

• Determine how much architectural information, such as hidden dimensions and embedding layers, can be extracted from production LLMs using query-based model stealing attacks.

Develop Efficient Extraction Techniques:

• Propose novel high-fidelity attacks that efficiently recover the embedding projection layer through targeted API queries.

Evaluate Practicality and Cost:

• Analyze the efficiency of the attack in real-world scenarios, measuring the cost and number of queries required to recover specific parameters.

Raise Awareness and Explore Defenses:

• Highlight the vulnerabilities in LLM APIs and propose potential mitigation strategies, such as rate limiting, noise addition, and architectural changes, to enhance model security.

In modern clinical practice, diagnosing illnesses involves analyzing diverse data like patient history, symptoms, and diagnostic reports, which can be time-consuming and error-prone. This project aims to develop a scalable and resolute MultiModal disease diagnosis and treatment system focused on cardiovascular function. Using ML, I integrated echocardiographic videos with a patient condition severity dataset to assist physicians in diagnosing and suggesting treatments. 

Objectives

Recent advances in biomedical sciences have boosted the usage of ML in assessing cardiac functions. Stanford's EchoNet-Dynamic outperforms human experts in segmenting the left ventricle (LV), estimating ejection fraction (EF), and assessing cardiomyopathy.

Another study used a DL model to estimate Right Ventricular EF (RVEF) from 2D echocardiographic videos . Researchers also explored supervised ML algorithms to predict cardiovascular diseases using variables like age, gender, and weight.

Despite these advancements, a significant gap remains in research that focuses on developing MultiModal disease diagnosis and treatment plans, which this project aims to address.

Results

We were able to beat the Benchmark for the prediction where RMSE was around 9 we were able to make it down to 8. Improved prediction accuracy and interpretability by supplementing the RNN with clinical features from patient datasets. 

The project was developed to address the critical issue of limited medical access in rural and semi-urban regions of India, where language barriers often hinder effective communication and diagnosis. This project aimed to provide an accessible, Hindi-language medical diagnosis system leveraging Large Language Models (LLMs). By receiving symptoms in Hindi, predicting diagnoses, and delivering results in the same language, the system empowers non-English speaking populations to access timely medical insights. 

Objectives

The objective of this project is to develop a scalable and accessible medical diagnosis system tailored for rural and semi-urban regions of India, where limited medical access and language barriers hinder effective healthcare delivery. The system aims to accept user-inputted medical symptoms in Hindi, translate them to English for processing by an ensemble of pre-trained Large Language Models (LLMs), and then provide a reliable diagnosis translated back into Hindi, preserving the input format (JSON or paragraph). By overcoming the challenge of English-dominated training data in existing LLMs, the project seeks to bridge the healthcare gap for Hindi-speaking populations, offering an inclusive, reliable, and user-friendly diagnostic solution. 

Results

By leveraging mBART and mT5 for translation, the system effectively converted user inputs from Hindi (local language) into English while preserving input formats such as JSON or paragraph. The translated English text was processed using an ensemble of domain-trained LLMs—LLaMA, PHI, and Mistral—to generate differential diagnoses and predict the most likely disease with high accuracy. A voting mechanism ensured reliability by combining outputs from the three models. The results were then back-translated into Hindi using mBART or mT5, maintaining clarity and consistency in the output. 

The project titled "Indoor Navigation for Visually Impaired" was designed to create a real-time assistive system that enhances indoor mobility for visually impaired individuals. This project was pursued to address the lack of accessible indoor navigation solutions, especially in complex environments like airports, museums, or offices. Developed at [Institution/Location], the project involved significant collaboration with peers and mentors, who provided valuable insights and support throughout the process. 

Objectives

The primary objective was to develop a computer vision-based navigation system capable of:

1. Mapping indoor spaces using keypoint detection, edge detection, and semantic segmentation.

2. Providing voice-based instructions with real-time obstacle detection and alerts.

3. Creating a reliable and user-friendly system that enables visually impaired users to navigate complex indoor environments independently.

Key considerations included ensuring real-time processing, accuracy of navigation, and adaptability to various indoor layouts.

Results

The project successfully delivered a robust indoor mapping and navigation system, integrating computer vision techniques for accurate virtual layouts and real-time obstacle detection. The voice-based navigation feature enabled users to receive turn-by-turn instructions, improving mobility and independence. A notable achievement was the system’s ability to adapt to different indoor spaces while maintaining accuracy. This project not only demonstrated the potential of AI in assistive technology but also paved the way for future developments in accessible indoor navigation systems. 

Newsletters are a treasure trove of insights, but the sheer volume of emails often makes it impossible to keep up. To address this challenge, I designed and developed an AI-powered Personalization Tool that revolutionizes how users interact with their newsletters. By leveraging Large Language Models (LLMs), this tool fetches newsletters from users' emails, personalizes content based on their preferences, and enables them to ask questions for tailored insights. This project demonstrates my expertise in AI integration, full-stack development, and user-centric design, while solving a real-world problem of information overload.

Objectives

The goal of this project was to create a tool that:

• • Reduces Information Overload: Automatically fetch and organize newsletters from users' inboxes.

  • Personalizes Content: Use AI to summarize, filter, and tailor newsletters based on individual preferences.

  • Enhances Engagement: Allow users to interact with newsletters by asking questions and receiving personalized answers.

  • Improves Productivity: Save users time by delivering only the most relevant and actionable insights.

Results

• • Successfully Built and Deployed: Delivered a fully functional tool that integrates with email providers (Gmail, Outlook) and processes newsletters in real-time.

  • Improved User Experience: Reduced newsletter reading time by 50% through AI-powered summarization and personalization.

  • High User Engagement: Achieved 90% user satisfaction in beta testing, with users praising the tool's ease of use and actionable insights.

  • Scalable Architecture: Designed a robust backend capable of handling thousands of newsletters daily, deployed on AWS/GCP for scalability.

Technical Achievements:

• • Integrated LLMs (OpenAI GPT) for content personalization and Q&A functionality.

  • Built a secure email fetching system compliant with GDPR and other privacy regulations.

  • Developed an intuitive frontend dashboard using React.js for seamless user interaction.

SupportSage is an intelligent, AI-powered customer support automation tool that leverages multi-agent orchestration to deliver fast, accurate, and empathetic support. By orchestrating LLM-powered agents, SupportSage automates issue resolution, ensures quality assurance, and personalizes user outreach, demonstrating technical expertise in LLM integration and process automation. 

Objectives

The goal of this project was to create a tool that:

• • Automate customer support workflows, reducing response time and operational overhead.

  • Ensure response quality through AI-based review and QA mechanisms.

  • Enhance customer satisfaction with personalized, context-aware support.

  • Integrate flexibly with popular CRMs and external management tools.

Results

• • Deployed a CrewAI-based support flow handling 100+ tickets/day with >85% auto-resolution rate.

  • Reduced response time by 60% and improved user ratings through tone-aware QA.

  • Plug-and-play architecture integrates quickly with enterprise SaaS tools.

Technical Achievements:

• • Multi-agent LLM orchestration (CrewAI, LangChain), custom tool plugins, secure CRM integration, real-world live ticketing pipelines.

ContentCrafter empowers writers, marketers, and businesses with a Jupyter-based AI platform for high-quality, customizable content generation. Leveraging LLMs for structured, SEO-friendly article drafts, ContentCrafter delivers creativity at scale—demonstrating skill in prompt engineering and user-facing notebook applications. 

Objectives

The goal of this project was to create a tool that:

• • Accelerate blog/article/email creation with guided workflows.

  • Enable customized tone, structure, and length to fit diverse needs.

  • Offer interactive, notebook-based editing for rapid prototyping and iteration.

Results

• • Cut content production time by 65%, enabling non-expert users to generate polished first drafts effortlessly.

  • Seamless integration with OpenAI/HuggingFace models and advanced prompt libraries.

Technical Achievements:

• • Modular Jupyter notebook architecture, OpenAI API integration, customizable prompt flows, React dashboard prototype for future scalability.

ResumeSensei streamlines resume customization by automatically rewriting and organizing resumes to fit any job description. Using advanced LLMs and role-targeted prompting, the tool delivers recruiter-friendly, ATS-optimized resumes—exemplifying practical AI for job market success. 

Objectives

The goal of this project was to create a tool that:

• • Generate tailored, role-optimized resumes for each job application.

  • Ensure ATS compliance and highlight relevant skills/achievements dynamically.

  • Provide an intuitive interface for users to adjust and export outputs.

Results

• • Increased interview call rates by up to 3× for power users.

  • Processes job descriptions and delivers a personalized resume in seconds.

  • Positive feedback for clarity, customization, and ease of use.

Technical Achievements:

• • LLM-driven content generation pipeline, keyword extraction using NLP, one-click PDF export via browser dashboard.