Projects

An LLM-powered RAG system offering hallucination-free, document-grounded answers, outperforming standard LLMs like ChatGPT in precision and reliability.

Built document-specific vector stores using sentence embeddings, enabling semantically accurate retrieval from complex data for factual, context-aware response generation.

Deployed on Microsoft Azure, DocDynamo delivers concise summaries, key insights, FAQs, and YouTube recommendations—all strictly derived from the uploaded documents.

Supports all common file formats—PDF, DOCX, TXT, PPT, and ZIP—ensuring seamless multi-format ingestion and intelligent parsing for enterprise-scale document analysis.

Developed a multi-agent system combining OCR and LLaMA-3 to automate contextual evaluation and feedback generation for student answer sheets.

Implemented a high-accuracy OCR Pipeline Agent to extract handwritten text from scanned sheets, ensuring reliable input for LLM-based assessment.

Built and deployed a UI Agent on Microsoft Azure, offering a responsive interface for seamless interaction between evaluators and students.

Engineered a UDP-based Tracking Radar Simulation System enabling real-time communication across multiple PCs, where each receiver listens via unique port numbers to simulate radar data transmission with minimal latency. 

Processed and simulated radar trajectories using ECEF, ENU, and LLA coordinate formats, extracting Range, Azimuth, and Elevation parameters, with real-time graphical visualization enabled through the GLG Toolkit interface. 

Implemented a receiver module that performs real-time data simulation, while systematically logging communication events and storing transferred data in structured files for complete traceability, auditability, and post-analysis verification.

Developed for the Ministry of Defence, Government of India, under the Smart India Hackathon 2023

Designed and trained a convolutional neural network (CNN) using TensorFlow to perform multi-class terrain classification, enabling automated environmental recognition from high-resolution satellite and drone-captured imagery datasets. 

Employed advanced data preprocessing techniques including normalization, augmentation, and reshaping to improve model generalization, reduce overfitting, and ensure robustness across varying terrain textures and lighting conditions. 

Achieved strong validation accuracy by fine-tuning hyperparameters and applying batch optimization strategies, validating the model’s effectiveness in distinguishing between complex natural landforms like deserts, forests, water bodies, and rocky terrain.