Research & Projects

Project 1

Text2SQL: Natural Language to SQL Query Generation System

This project was developed to convert natural language questions into executable SQL queries using a Retrieval-Augmented Generation pipeline. The goal was to bridge the gap between non-technical users and structured databases by combining semantic retrieval, schema-aware context, and a fine-tuned CodeT5+ model. The system uses a Flask frontend, a FastAPI backend, FAISS-based vector search, and GPU-backed model inference to generate context-aware SQL from plain English questions.

Objectives

The project aimed to build an end-to-end Text-to-SQL system that could accurately interpret user questions and generate relevant SQL queries against database schemas. Key objectives included enabling natural-language querying, improving contextual understanding through RAG, retrieving similar examples and schemas from a FAISS vector store, and providing a clean interface where users could compare the question, generated SQL, schema, and related examples side by side. The architecture was also designed to support scalable inference through a local FastAPI service connected to a Google Colab GPU backend via ngrok.

Results

The solution delivered a working Text-to-SQL application with a modular backend and frontend, semantic retrieval over 9,715 vectors spanning 206 databases and 8,659 examples, and a CodeT5+ model fine-tuned on Spider for SQL generation. The repository reports 60%+ execution accuracy on the Spider test set, while also demonstrating practical experience in LLM application design, RAG pipelines, vector databases, API integration, and full-stack deployment. This project strengthened my hands-on skills in building AI systems that combine retrieval, model inference, and user-facing interfaces for real-world data access workflows.

Project 2

RideShare: Distributed Ride-Matching and Monitoring System

This project was developed to simulate a real-time ride-sharing platform using distributed database concepts and transactional matching logic. The goal was to build an end-to-end system that could generate synthetic ride requests, clean and process trip data, assign the nearest available driver, and monitor ride activity through a live dashboard. The project combines Python-based data pipelines, CockroachDB, and a Flask web interface to demonstrate scalable ride allocation and analytics in a multi-node environment.

Objectives

The project aimed to build a ride-sharing workflow on top of a local 3-node CockroachDB cluster, with a strong focus on data processing, transaction safety, and real-time visibility. Core objectives included generating synthetic rides into a staging table, cleaning them into structured ride data, initializing drivers, replaying rides in sequence, and assigning the closest available driver using serializable transactions. Another key goal was to expose driver status, ride progress, and system analytics through a Flask and HTML dashboard for live monitoring.

Results

The solution delivered a working ride-matching system with modular components for database connection management, data loading, cleaning, driver initialization, ride replay, matching, and dashboard serving. The repository includes a structured pipeline with separate modules for synthetic data generation, transaction-based matching logic, and a frontend interface for observing assigned and unassigned drivers in real time. This project strengthened my hands-on experience in distributed systems, database transactions, backend development, and building operational dashboards for live system analytics.

Project 3

SprintTrack: AI-Powered Jira Assistant

This project was developed as an intelligent productivity tool to help Agile teams manage Jira workflows more effectively. The goal was to combine Jira data with AI capabilities in a Streamlit-based dashboard so users could generate sprint summaries, ask natural-language questions, refine backlogs, create release notes, and automate project-analysis tasks from a single interface. The system integrates Jira APIs and OpenAI-powered workflows to improve visibility, reduce manual effort, and support faster decision-making for teams.

Objectives

The project aimed to build an AI assistant for Jira that could simplify sprint management and backlog analysis through automation and retrieval-based intelligence. Key objectives included generating role-based sprint summaries, enabling natural-language Q&A over Jira issues using RAG, assisting with backlog refinement through acceptance criteria and estimation suggestions, and supporting engineering workflows with release-note generation, test-case creation, ticket triage, and risk analysis. The broader focus was to make Agile reporting and planning faster, more contextual, and more actionable.

Results

The solution produced a functional Streamlit dashboard with multiple AI-assisted Agile features, including sprint summaries, backlog refinement, release notes, retrospectives, ticket labeling, and onboarding documentation. The repository is structured with modular components for the Streamlit app, Jira API integration, and issue-data processing, making the system extensible for additional team workflows. This project strengthened my hands-on experience in building AI applications with Python, integrating external APIs, designing RAG-style workflows, and developing practical tools for real-world software delivery teams.

Naturallangprocessors-poster.pdf

Project 4

Translated Navigation Instructions for Robot Navigation

This project was conducted during my undergraduate studies at IIT Gandhinagar as part of a natural language processing course. The goal was to bridge human language and robotic execution by converting navigation instructions into actionable commands for autonomous systems.

Objectives

The project aimed to build an NLP pipeline that could accurately parse and translate complex human instructions into robot-compatible language. Major challenges included dealing with unstructured language, ensuring contextual accuracy, and optimizing performance on large datasets

Results

The solution achieved 82% accuracy on a gigabyte-scale dataset by implementing a "learning in chunks" strategy, which outperformed baseline models. This reinforced my interest in NLP applications and gave me valuable hands-on experience in data preprocessing, tokenization, and model training.

Project 5

Advanced Multimedia and Web Database Systems: Human Motion Video Analytics

This project was completed as part of an advanced database systems course at Arizona State University. The goal was to design a multimedia video retrieval system leveraging computer vision and deep learning techniques. I worked independently with guidance from course instructors to apply theory in a practical and scalable system.

Objectives

The primary aim was to implement a video analytics system that could classify and retrieve human motion activities from a large video dataset. Key objectives included accurate feature extraction using deep learning (R3D-18), clustering for organization, and dimensionality reduction for performance optimization. Considerations included system accuracy, search speed, and similarity ranking quality.

Results

Successfully built a prototype that achieved 77% retrieval accuracy using Locality Sensitive Hashing. Classification of 51 motion types was achieved through k-NN and clustering with K-means. Advanced reduction techniques like PCA, SVD, and LDA improved efficiency, confirming the system’s potential for real-world video search applications.

Project 6

Parameter Optimization using Machine Learning in Additive Manufacturing | Skills Gained: ANN

Conducted at IIT Gandhinagar, this project explored the use of artificial neural networks (ANN) to optimize parameters in Selective Laser Melting (SLM), a key process in metal additive manufacturing. The project blended materials science with machine learning, aligning with my interdisciplinary academic background.

Objectives

The goal was to develop a predictive model that could optimize over five critical process parameters to improve part quality (e.g., density, microstructure). This involved creating an ANN capable of understanding the relationship between input variables and quality outcomes in Powder Bed Fusion techniques.

Results

The ANN model successfully identified optimal parameter ranges, leading to improved prediction accuracy for key quality metrics. The project demonstrated the value of ML in industrial applications and sparked my deeper interest in intelligent manufacturing systems.

Page updated

Google Sites

Report abuse