ML/AI Engineer & Researcher

Built a Retrieval-Augmented Generation pipeline with a fine-tuned 4-bit quantized LLaMA-2 model using Low-Rank Adaptation (LoRA). Integrated FAISS vector database for efficient document retrieval and similarity search. Optimized memory and retrieval latency by combining dynamic batching, cosine learning rate scheduling, and quantization-aware inference.

Built a real-time fraud detection system using Apache Kafka, PySpark Structured Streaming, and FastAPI, capable of identifying fraudulent transactions with over 95% precision on an imbalanced dataset. Trained a class-imbalance-aware XGBoost model and deployed it as a lightweight microservice using Docker, achieving sub-50ms prediction latency. Scalable up to 1M+ transactions/day.

Developed a Reinforcement Learning agent using Deep Q-Network (DQN) architecture to play and improve performance in a custom-built Snake game environment. Designed and trained a Neural network model with checkpointing, experience replay and fine-tuning strategies, enabling the agent to achieve superhuman performance with an average high score of 150 across 120 games.

Built and deployed an interactive Streamlit web application to predict obesity risk based on 14 health and lifestyle features using Machine Learning models with dynamic hyperparameter tuning. Enhanced model explainability with LIME, SHAP, and PDPBox to visualize feature impact and provide transparent health risk insights for end users.

Interactive data visualization dashboard built with Streamlit for exploratory data analysis and ML model deployment. Features comprehensive data analysis tools, machine learning model integration, and dynamic visualization capabilities using Pandas, XGBoost, and Seaborn for professional-grade data insights.

A web-based to-do list application created using Streamlit and SQLite3. Features a clean, intuitive interface for task management with persistent storage using SQLite database. Implements CRUD operations for managing tasks efficiently with real-time updates and responsive design for seamless user experience.

A comprehensive placement preparation portal built with Python-Flask backend, featuring real-time communication via sockets, RESTful API integration, and responsive frontend using HTML, CSS, and JavaScript. Includes YAML-based configuration management and structured data handling for student placement resources and company information.

Full-stack flashcard application built with Python-Flask backend and Vue.js frontend. Features asynchronous task processing with Celery, YAML-based configuration, and a modern responsive interface. Implements spaced repetition algorithms and user authentication for personalized learning experiences.

Comprehensive data analysis project using Jupyter Notebook to analyze optical sales and customer data. Implements statistical analysis, data visualization, and predictive modeling techniques to derive actionable business insights. Features exploratory data analysis, trend identification, and customer segmentation using Python data science stack.

Implementation of Conway's Game of Life cellular automaton using Python and PyGame. Features interactive grid manipulation, multiple pattern presets, adjustable simulation speed, and real-time visualization. Demonstrates computational thinking and algorithmic problem-solving in simulating complex emergent behaviors from simple rules.

Voice-activated virtual assistant built with Python using speech recognition and text-to-speech capabilities. Features include web browsing automation, system command execution, information retrieval, and natural language processing. Integrates multiple APIs for weather, news, and general queries with hands-free voice control.

Browser extension calculator built with vanilla HTML, CSS, and JavaScript. Features a clean, responsive interface with support for basic arithmetic operations, keyboard shortcuts, and memory functions. Demonstrates proficiency in browser extension development and front-end engineering fundamentals.