BihariBlogger – A Social Blogging Web App

Overview:
BihariBlogger is a dynamic blogging platform that allows users to share their thoughts, connect through posts, and explore content from fellow bloggers. It is designed with scalability, usability, and modularity in mind using the Flask web framework and MongoDB for flexible document storage.

Key Features:

• User authentication with registration and login system using Flask-Login.
• Clean and responsive UI built with HTML5, CSS3, and Bootstrap 5.
• Write, edit, and delete blog posts with Markdown-style content support.
• Contact and feedback form integration with email alerts (SMTP setup planned).
• Each user has their own dashboard with a personalized list of authored blogs.
• MongoDB backend for flexible schema design and fast CRUD operations.

Technical Stack:

• Backend: Flask, Flask-PyMongo, Flask-WTF, Flask-Login
• Frontend: HTML, CSS, Bootstrap 5, Jinja2 templating
• Database: MongoDB (NoSQL) via PyMongo
• Additional Tools: Jinja2, WTForms for form validation, Flask’s `url_for` and Blueprints for modular routing

Challenges Overcome:

• Handled CSRF protection and form validation using Flask-WTF.
• Designed RESTful routes and maintained session-based authentication with Flask-Login.
• Built a modular file structure separating routes, forms, and templates for clean scalability.

Future Improvements:

• Implement user avatars and profile bios
• Allow liking and commenting on posts
• Add category filtering and search functionality
• Enable email notifications and password reset flows

GitHub: View Repository

MockAnalysis

The MockAnalysisProject is a smart automation tool built to extract, clean, and analyze students' mock test performance data from an online educational platform. This project was designed to automate the repetitive process of manual result tracking and enable insightful analysis for academic performance monitoring.

The system uses Selenium to programmatically interact with the test portal, including logging in, navigating result pages, and extracting structured data. Once collected, the data is cleaned and processed using Pandas for tabular transformation and aggregation.

The tool is highly customizable and can be adapted to work with different institutional platforms that follow structured formats.

GitHub: View Repository

Gender Classification using CNN

This project focuses on building a deep learning-based model that classifies a person's gender from facial images. It utilizes Convolutional Neural Networks (CNNs) to learn intricate facial features for high-accuracy predictions.

A dataset of over 3,000+ labeled facial images was cleaned and preprocessed (resized, normalized, and augmented) before being fed into the CNN. The model was trained and fine-tuned to detect and classify gender with over 87.8% accuracy.

Real-time face detection was integrated using OpenCV, enabling live predictions directly from a webcam. This significantly enhances the model’s practical usability for real-world applications like demographic analysis and personalization engines.

Additional improvements were achieved by experimenting with different kernel sizes, dropout layers, and batch normalization to reduce overfitting and improve generalization.

GitHub: View Repository

Heart Disease Prediction

A machine learning-based diagnostic tool that predicts the presence of heart disease using clinical data such as age, sex, chest pain type, resting blood pressure, cholesterol levels, and other health indicators.

• Performed data preprocessing, feature selection, and model training using Logistic Regression, Random Forest, and Support Vector Machines.
• Implemented extensive exploratory data analysis (EDA) and understand correlations and feature importance.
• Evaluated models using confusion matrix, ROC-AUC, precision, recall, and F1-score, achieving an accuracy of ~86%.
• Deployed the application using Streamlit for an intuitive and interactive web interface.

GitHub: View Repository
Live App: Launch on Streamlit

Smart Traffic Management System (AI + MLOps)

Overview:
End-to-End Computer Vision + MLOps Pipeline using YOLO, Airflow, MLflow & DVC A production-oriented smart traffic analytics system that performs vehicle detection, tracking, counting, and traffic flow analysis, orchestrated using Apache Airflow and fully versioned with DVC + MLflow.

Key Features:

• Vehicle Detection: YOLO-based real-time object detection.
• Object Tracking: Persistent ID tracking across frames.
• Traffic Metrics: Vehicle count, flow, density estimation.
• Dataset Validation: Integrity checks & auto-labeling.
• Experiment Tracking: MLflow logging & reproducibility.
• Visualization: Streamlit dashboards

Tech Stack:

Key Observations:

• Car detection achieved excellent performance, making it highly reliable for vehicle counting and traffic flow estimation.
• Bus and truck classes show strong detection accuracy, suitable for heavy-vehicle analytics.
• Person detection performs well, enabling pedestrian density and safety analysis.
• Motorbike detection shows moderate performance, mainly affected by small object size and occlusion.
• When combined with multi-object tracking (ByteTrack / BoT-SORT), the effective accuracy improves significantly, ensuring stable real-world deployment.

Future Improvements:

• Multi-camera tracking.
• Detect whether the driver is wearing a helmet and notify administrators via IoT (Raspberry Pi).
• Vehicle re-identification.
• Number plate recognition.
• Detection of traffic rule violations.
• Real-time Kafka-based data ingestion.

GitHub: View Repository

PIMA Diabetes Prediction

Built a predictive ML tool using the PIMA Indian Diabetes dataset to identify the likelihood of diabetes in female patients based on medical and lifestyle data such as glucose level, BMI, insulin level, and age.

• Utilized Logistic Regression, K-Nearest Neighbors, Decision Trees, and ensemble techniques for improved model performance.
• Applied data normalization and feature scaling to handle wide-ranging feature values.
• Conducted EDA and plotted histograms, pair plots, and correlation matrices for feature insights.
• Achieved an accuracy of ~82% and deployed it as an interactive web app using Streamlit.

GitHub: View Repository

Stock Market ETL Pipeline (Apache Airflow + MongoDB)

This real-world ETL pipeline automates the daily ingestion of stock price data from an external API. Designed for production-readiness, it includes fault-tolerant architecture, task modularity, and streamlined scheduling using Apache Airflow.

• Data Flow: Extract → Transform → Load, fully orchestrated via Airflow DAG
• Transformation: Data cleaned and processed using Pandas
• Storage: Structured stock data stored in MongoDB as JSON
• Configurable: Task retries, XComs, logging, and scheduling support

Tech Stack: Apache Airflow, Python, Pandas, MongoDB, Streamlit (dashboard in progress)

What I Learned: Real-world data workflow design, modular ETL architecture, NoSQL integration, and Airflow DAG best practices.

Planned Improvements:

• Slack/email notifications for task failures
• Streamlit dashboard for visualizing top gainers/losers
• Dockerized deployment for portability
• Real-time stock prediction using RNN models

GitHub: View Repository

Next Project in Progress