PROJECTS

Paste a YouTube URL and instantly chat with the video, generate structured summaries, or explore the transcript with keyword search and timestamp navigation. Built with LangChain, Groq LLaMA 3.3, FAISS vector search, and HuggingFace embeddings — all running on Streamlit. Supports multi-session chat, multilingual transcripts with English responses, and one-click export of chats and summaries.

Upload one or more PDFs and have a real conversation with them — not just one-off questions, but proper back-and-forth with memory. Built with LangChain, Groq LLaMA 3.3, ChromaDB, and HuggingFace embeddings. Handles follow-up questions intelligently by reformulating them into standalone queries before retrieval, so context never gets lost across turns.

A production-grade MLOps pipeline for 3-class Twitter sentiment classification (Positive / Negative / Neutral). Covers the full lifecycle — data ingestion from S3, NLP cleaning, CountVectorizer feature engineering, LightGBM training with Optuna hyperparameter tuning, experiment tracking on DagsHub via MLflow, and automated model promotion from staging to production. Exposed as a Flask REST API, containerized with Docker, and deployed to AWS EC2 via CodeDeploy with a full GitHub Actions CI/CD pipeline enforcing accuracy, F1, precision, and recall quality gates before every deployment.

End-to-end MLOps pipeline for predicting food delivery time using a stacking ensemble of CatBoost and Random Forest with a Decision Tree meta-model. Built with DVC for pipeline reproducibility, MLflow on DagsHub for experiment tracking, and a structured model promotion flow — candidate → staging → production — with R², MAE, latency, and extreme error gates before every promotion. Served via a FastAPI inference service containerized with Docker, deployed to AWS EC2 through CodeDeploy with full GitHub Actions CI/CD automation.

Built an end-to-end number plate detection system using deep learning. Detected license plates using a YOLOv8m model trained on the Indian vehicle dataset  and optimized for accurate, real-time detection. Developed a Gradio interface and deployed on Hugging Face to enable live image and video-based predictions.

Classifies IMDB movie reviews as positive or negative using a SimpleRNN trained on 50,000 reviews with Word Embeddings for sequence modeling. Trained with EarlyStopping, ReduceLROnPlateau, and ModelCheckpoint callbacks, achieving 82% test accuracy. Deployed as an interactive Streamlit app where you can paste any review and get an instant sentiment prediction with a confidence score.

A personalized movie recommendation system built using content-based filtering. By analyzing features like genres, keywords, cast, and movie overviews, it identifies patterns in your preferences and recommends similar titles. Powered by vector space modeling and cosine similarity, the system offers intelligent recommendations straight from the TMDB 5000 Movie Dataset.

An end-to-end data-driven web application designed to make property buying decisions smarter. Built with Streamlit, it integrates interactive visualizations—including sector-level geomaps, price distribution charts, and feature word clouds—with a machine learning–powered price prediction model and a content-based recommender system that suggests similar properties based on user preferences. Users can explore market trends, compare sectors, and get personalized property recommendations, making it a valuable tool for buyers, sellers, and real estate analysts.

Built an end-to-end face emotion recognition system using deep learning. Faces are detected with MTCNN and classified into 7 emotions using a ResNet50 model with transfer learning. Implemented MLflow for experiment tracking and deployed via a Gradio app on Hugging Face for real-time predictions.

Predicts whether a bank customer will churn using an ANN trained on 10,000 records, achieving 86% accuracy and 0.86 ROC-AUC. The network uses two hidden layers with ReLU activation, trained on features like credit score, geography, age, and account balance with label encoding, one-hot encoding, and standard scaling applied in the preprocessing pipeline. Deployed as a Streamlit app where you input customer details and get an instant churn probability with a clear yes/no decision.