Projects
‍

AI-Powered Fashion Agent involved in by leveraging the techniques of generative ai to designing three new innovative outfit sets (top wear, bottom wear, footwear, accessories) and shop products online, based on user-uploaded images, style preferences, budget, and occasion by using Agentic workflow with crewai. Project results turns into high impact in improving user experience and reduced lot of user online product searching time. Each outfit is presented in a detailed blog with complete details of the whole outfit set(summary, title, product prices, links, images and each outfit set price.)

RAG based AI chatbot involved in by leveraging the techniques of generative ai and built AI-powered search and chatbot interaction for exploring Renesas electronics device designs. It crawls the Renesas website, extracts device descriptions and block diagrams, and organizes them for intelligent retrieval. Using BeautifulSoup, Hugging Face embeddings, and a FAISS vector database, it supports fast similarity search. It enables users to input queries, view the top two matching devices, and engage with a chatbot powered by LLM to ask questions about the retrieved devices and streamline technical research and design analysis for engineers.

Rice project is an end-to-end Machine learning classification project, which demonstrate mainly on the life cycle of Machine learning project and AWS cloud deployment with principles of CI/CD pipeline and AWS services(SageMaker, lambda, CodeBuild, CodeDeploy, CodePipeline, CodeCommit, s3 bucket, ECR, EC2, IAM). Implemented the data preprocessing job by setting up the SageMaker processing container and saved the preprocessed data in s3 bucket. With the final trained model created the endpoint in SageMaker for future data prediction.

This project delivers an AI-powered recommendation system for Amazon apparel product, combining text and image similarity to suggest related fashion items. It processes 180,000+ products by cleaning metadata, applying Python and state-of-the-art NLP and deep learning libraries, BoW, TF-IDF, Word2Vec, and VGG-16-based visual embeddings for multi-modal analysis. A hybrid similarity approach incorporates brand and color metadata for more accurate recommendations. The system enhances product discovery and personalization, supporting e-commerce use cases like user engagement, inventory management, and marketing.

This project delivers an AI-powered solution to automatically classify real estate sales messages as <lead>, <ignore>, or <spam>, enhancing efficiency for a real estate-based company. It uses Meta’s LLaMA-3.1–8B model fine-tuned with domain-specific conversations using LoRA and 4-bit quantization for optimized performance. The system processes dialogues with custom prompts, performs scalable inference via a Hugging Face-hosted model, and achieves high contextual accuracy. This reduces manual triage, filters spam, and streamlines the sales pipeline with a scalable, cost-effective LLM-based solution.

Document Classification involved in by leveraging the techniques of deep learning neural networks and NLP techniques to build an end-to-end text classification system to categorize 18,000+ documents into 20 classes using deep learning. It features advanced preprocessing—extracting subject lines, cleaning metadata, applying NLP techniques, and handling noisy text. Two 1D CNN models were implemented: one with GloVe embeddings and another with character-level embeddings, optimized using developed custom callbacks, early stopping and evaluated via F1 score and accuracy. The solution enables scalable, high-accuracy document classification with real-world applications like email filtering and legal text analysis.

This project addresses Quora’s challenge of detecting duplicate questions to improve content consolidation and user experience. Framed as a binary classification task, it combines NLP preprocessing, hand-engineered features, and TF-IDF vectors to model question similarity. Techniques like fuzzy matching, token overlap, and t-SNE visualization enhance feature extraction, while models like Logistic Regression, SVM, and XGBoost are used for prediction. The system enables scalable, accurate duplicate detection, reducing content redundancy and streamlining knowledge discovery.