Agentic Voice Assistant with Task Management and Search Capabilities

Developed an AI-powered voice assistant capable of performing complex multi-step tasks such as setting alarms, creating Google Calendar events, sending emails, and conducting internet searches.

Incorporated real-time voice input with the SpeechRecognition library for dynamic task execution and user interaction.

Used SerpAPI for Google search automation, Plyer for desktop notifications, and smtplib for seamless email handling.

Knowledge Graph Q&A and RAG with Tabular Data

Architected a chatbot for retrieving insights and answering user queries based on structured tabular data (CSV, XLSX) and unstructured text.

Constructed a dynamic knowledge graph using Neo4j to represent complex relationships between entities for enhanced data retrieval.

Integrated LangChain and Neo4j graph database to enable efficient vector-based semantic search and contextual query resolution.

Designed and implemented Cypher queries for optimized graph traversal, ensuring accurate and relevant data retrieval.

Enabled real-time fuzzy and Soundex search functionalities to improve user experience and query accuracy.

Resume Changes Suggestion Chrome Extension

Engineered a Chrome extension leveraging GPT-4 to analyze resumes against job descriptions and offer tailored optimization suggestions.

Integrated LangChain with Pinecone Vector DB to store and retrieve job descriptions, enabling contextual and precise resume enhancements.

Enabled real-time job description extraction from websites, allowing seamless interaction and instant feedback.

Designed an intuitive UI for effortless resume analysis and actionable suggestions, ensuring user engagement.

Reflexion Agent with LangGraph

Programmed a Mental health assistant Reflexion agent for structured reasoning and dynamic memory retrieval, used iterative learning and decision-making from user interactions.

Implemented an adaptive Reflexion agent using LangGraph for structured reasoning and dynamic memory retrieval, enabling iterative learning and decision-making from user interactions.

Integrated Pinecone vector database to provide context-aware responses and maintain a memory of prior interactions, enhancing the agent's ability to precisely handle complex, multi-turn conversations.

Leveraged GPT-4 for natural language understanding, enabling the agent to refine actions and improve decision accuracy through reflective reasoning and knowledge updates.

Hindi Chatbot for Law and Medical Assistance

Engineered a Hindi chatbot for Law and Medical assistance by integrating a centralized Ollama LLM with LoRA fine-tuned models

Utilized OpenAI MMMLU Hindi datasets to deliver accurate domain-specific query responses

Fine-tuned domain-specific LLMs using PyTorch for precise and context-aware responses tailored to the nuances of Hindi language usage and Unsloth to Load and Save the models to HuggingFace.

Cold Email Generation System

Build a system to Web scrape multiple websites for the opportunities and generate a cold email response using LangChain, Prompt engineering, and RAG with Pinecone Vector DB to store and retrieve vectorized resume data.

Integrated GPT-4 to generate personalized email drafts, ensuring relevance and effectiveness, with Python handling the backend integration

Developed a simple UI for users to review and edit generated emails before sending, integrating this with the backend for seamless operations.

Fraud Detection System

Implemented a fraud detection system leveraging LLMs and RAG to analyze financial statements in PDF, detect anomalies, and generate detailed fraud reports using tools like Chroma vector store, Hugging Face, NLTK, and Pandas for efficient processing and retrieval.

Integrated Chroma vector store for efficient management of vectorized financial data, enabling fast retrieval and similarity searches to detect patterns of fraud by comparing current statements with historical ones.

Leveraged NLTK for text preprocessing, improving data quality by performing tokenization, stopword removal, and lemmatization to prepare the data for anomaly detection.

Custom Summarization Application with T5

Fine-tuned the T5 transformer model using PyTorch and Hugging Face for a domain-specific summarization task, optimizing performance through tailored training datasets and hyperparameter tuning to generate concise and contextually accurate summaries.

Created and preprocessed custom training datasets that reflected the unique characteristics of the domain, ensuring that the T5 model was exposed to relevant data for improved performance in summarizing specialized content.

T-Shirt Store Inventory RAG Application

Created a RAG system for a T-shirt store, allowing users to retrieve product details, inventory status, and pricing directly from a MySQL database using natural language inputs, with GPT-4 generating and executing SQL queries to enhance user engagement and operational efficiency.

Configured GPT-4 to generate and execute SQL queries based on user inputs, allowing users to ask questions like "What is the inventory status for size M?" and receive immediate, accurate responses directly from the database.

Designed a user-friendly interface that allows staff to interact with the system via natural language, simplifying the process of checking inventory levels, product descriptions, and pricing without requiring SQL knowledge.

LLM-Assisted Compatibility Test App

Built an LLM-powered application using Google’s Gemini API and Django to analyze Android app screenshots and determine login success.

Integrated AWS services (EC2, S3) and PostgreSQL for secure and efficient backend operations.

Developed a responsive frontend with React.js, enhancing user experience and implementing robust security measures.

Delivered a scalable and cloud-based solution, showcasing expertise in full-stack development and AI-driven automation.

B2B Agriculture Platform

Built a B2B e-commerce farmer-focused web application platform for trading agricultural commodities, integrating MVC architecture, Spring Boot, and ReactJS with 7 backend services for product listings, authentication, and transactions.

Built 7 backend APIs including product listings, user authentication, transaction processing, order management, and inventory tracking, enabling a seamless trading experience for farmers and buyers

Utilized Spring Boot for building RESTful APIs, ensuring smooth communication between the frontend and backend, and allowing for easy integration with external systems.

Institute ERP System

Created an Educational ERP system with five administrative modules to automate student admissions, attendance tracking, grade management, fee collection, and payroll scalability and seamless operation using ReactJS, Java, and Oracle.

Designed and implemented a user-friendly ReactJS interface for administrators, teachers, and students, offering intuitive navigation and real-time updates for student-related activities.

Implemented a fee collection system, allowing students to pay fees online, track payment history, and receive notifications for upcoming fee deadlines, improving financial transparency and reducing administrative overhead.

Face Emotion Detection

Designed a deep learning algorithm for emotion detection using FER 2013 and CK+ datasets, achieving enhanced accuracy with selective learning and publishing findings at the IEEE Conference (2021).

Utilized TensorFlow and Keras for model implementation, optimizing network architecture and training processes to increase the system’s accuracy and real-time performance.

Preprocessed and augmented large image datasets (over 10,000 images), optimizing model training with techniques like image normalization and data augmentation to improve model robustness and performance

Document Summarization System

Implemented a comprehensive document precise summarization system for efficient retrieval, and contextual understanding of extensive and complex documents for diverse use cases.

Integrated Pinecone Vector DB to store document embeddings and enable fast, scalable retrieval of relevant information from vast document datasets, ensuring efficient summarization of lengthy or complex documents.

Implemented Retrieval-Augmented Generation (RAG) architecture to combine the strengths of both information retrieval and generative models, improving the relevance and accuracy of summaries generated from a large corpus of documents.