• Led research on a novel framework for RL safety assurance, combining local and global policy explanations into a unified directed graph abstraction to enable interpretable and structured safety analysis.
• Applied formal verification with the Storm model checker on the abstracted policy graph and integrated a risk- and uncertainty-guided falsification strategy to explore high-risk or underrepresented states.
• Developed an ensemble of deep neural network encoders, trained via contrastive learning, to guide the agent toward high-risk and uncertain regions to uncover violations missed by abstraction.
• Achieved 30% more safety violation detections than baseline methods across three high-stakes domains: Navigation2, Maze (CMDB Mujoco), and a simulated insulin dosing task for Type 1 Diabetes.

• Researched and implemented approximate unlearning methods to reduce toxicity in large language models by removing memorized toxic data subsets, achieving 15% reduction in toxicity metrics.
• Fine‑tuned multiple LLMs using toxic prompt templates and datasets and evaluated their toxicity reduction post‑unlearning, proving improved ethical model behavior.

• Developed a fast and sample-efficient approach for multi-task active learning when the amount of data from source tasks and target tasks is limited.
• Proposed an adaptive sampling-based alternating projected gradient descent (GD) and minimization algorithm that iteratively estimates the relevance of each source task to the target task.
• Evaluated proposed algorithms on Corrupted MNIST datasets, achieving over 12% greater efficiency compared to baseline multi-task learning approaches.

• Developed an emotion recognition system using Dynamic Routing for Vision Transformer and Convolutional Neural Networks to recommend music based on human facial expression.
• Achieved a 9.7% improvement in the performance of the emotion recognition system on the AffectNet dataset, enhancing model accuracy across diverse facial orientations.
• Integrated the emotion recognition system into a music recommendation platform, demonstrating practical application of AI in enhancing user experience through personalized content suggestions.

• Designed and developed a group equivariant version of the PointNet-like Multi-Layer Perceptron for the 3D point cloud objects, contributing to theoretical proving its efficacy as a group-equivariant universal approximator.
• Implemented the innovative G-PointConv group equivariant Convolutional Neural Network, achieving state-of-the-art performance in 3D point cloud classification tasks (91.77% on the original ModelNet40 dataset; 91.21% on random rotated datasets).

• Designed AI Agent using AWS Bedrock and Claude LLMs to automate SQL query optimization, reducing manual database administration tasks by 85% and improving query performance across enterprise systems.
• Built scalable RAG (Retrieval-Augmented Generation) system indexing 2000+ enterprise database tables to provide context-aware SQL recommendations and automated troubleshooting.
• Integrated real-time monitoring capabilities to track query performance metrics and automatically suggest optimization strategies based on historical patterns and database usage analytics.

• Analyzed 3M customer service calls using advanced statistical methods and machine learning techniques to identify patterns, trends, and optimization opportunities in healthcare revenue cycle management.
• Developed and deployed machine learning models using XGBoost and CatBoost algorithms on AWS SageMaker, achieving 93.11% accuracy in call volume forecasting to optimize staffing and resource allocation.
• Created automated data pipelines and interactive dashboards to provide real-time insights for business stakeholders, enabling data-driven decision making and improving operational efficiency.

A full-stack AI chatbot built with Next.js that can interact with users in real-time using streaming responses. The agent can execute various tools autonomously and provides a terminal-like interface to display tool operations, making it suitable for a wide range of applications including data retrieval, analysis, and task automation.

• Real-Time Streaming Chat Interface: Implemented Server-Sent Events (SSE) for streaming responses with typing indicators, providing immediate feedback and natural conversation flow using Next.js API Routes and React components.
• Autonomous Tool Execution Framework: Integrated LangChain and LangGraph for agent workflow orchestration with Claude 3.5 Sonnet, enabling the AI to autonomously select and execute appropriate tools based on user queries with visual terminal-like display.
• Multi-Modal Information Retrieval: Built comprehensive tool suite including Google Books API integration for book search and detailed information retrieval, Wikipedia Search API for encyclopedia content, and YouTube transcript extraction supporting multiple languages.
• Enterprise-Grade Authentication & Data Management: Implemented Clerk authentication system with Convex database for persistent chat history, user management, and secure data handling across multi-turn conversations with context preservation.
• Advanced Customer Data Analytics: Developed customer information retrieval system with access to names, addresses, order histories, shipping tracking, and comments data including user engagement metrics and post analytics.

SQL Query Generator Using Large Language Model. Through a user-friendly web application, users input database schemas, feature specifications, and requirements in natural language. SQLGenius will generate the SQL query for developers and non-technical users to get insights from data.

• Efficient Database Schema Management: Users can define and edit database schema by interacting with the UI or uploading CSV files containing the data, streamlining the schema creation process.
• Database Schema Visualization: Provides visual representation of the input database schema including table structures and column relationships for better understanding and validation.
• Natural Language SQL Generation: Converts natural language requirements into SQLite code by gathering information from database schema and user specifications, bridging the gap between technical and non-technical users.
• Automated Query Creation: Significantly accelerates query development process, offering an efficient and accessible solution that revolutionizes database interaction for both developers and non-technical users.

A next-generation movie streaming platform offering a feature-rich viewing experience, equipped with real-time interactive capabilities for movie enthusiasts and cinephiles.

• Secure User Authentication: Implemented seamless sign-in using Google and traditional email/password methods with Firebase integration, ensuring secure data handling and smooth user access.
• Extensive Movie Library: Built comprehensive movie database integration with The Movie Database (TMDB) API, enabling users to explore vast collections and stream both classic favorites and new releases.
• Real-Time Commenting System: Developed community-driven commenting functionality using Socket.io that updates in real-time, allowing users to share thoughts and insights on scenes as they unfold during viewing.
• Personalized Recommendations: Created intelligent recommendation engine that learns user viewing habits and preferences to provide tailored movie suggestions, enhancing content discovery.
• High-Quality Streaming: Implemented adaptive streaming technology that automatically adjusts video quality based on internet speed, ensuring optimal viewing experience across different connection types.