Tejal Rawale

Call Selling Price Prediction

An end-to-end Machine Learning project that helps users to get an estimated selling price for their used car based on various features, which can help in making informed decisions when selling a car.

Data Source:
Data is collected from Car Dekho portal (a website to buy new/used cars in India).

Data Processing:
Performed well thorough EDA including data cleaning, pre-processing (check for missing values, imputations, invalid values, outliers), statistical analysis, analysis to identify trends and patterns in the data through visualization.

Machine Learning :
Further data partitioning was performed to train-validate-test the data on the model. Feature engineering (normalization, encoding, feature selection) was added to the machine learning pipeline. Since the aim of the project is to predict the reselling price of the car (continuous variable), it is a regression task. I tried different Regressor ML algorithms and evalauted their performance using regressor metrics. The Random Forest regressor proved to have best performance after evaluation. This model was trained on the entire data to make it ready for deployment.

Front-end :
I designed a simple yet intuitive user-friendly UI for the users to interact with the model. The front end interface was developed using HTML, CSS, Javascript within Flask.

Deployment :
The trained regressor model was deployed using Flask web application.

Tech Stack :
python, pandas, numpy, seaborn, mathplotlib, scikit-learn, flask, API, javascript, html, css.

Movie Review Analysis

Customer centric businesses heavily rely on product reviews to improve their​ product and maximize their market value. In this project we focus to perform sentiment analysis on the reviews to gauge the overall positive or negative valence of the review and then perform topic modeling to see what topics are prominent from the reviews for a particular movie.

Data Sources:
1. IMDb 50K movie reviews dataset released by Stanford.
2. Reviews of The Shawshank Redemption scraped from IMDb using Selenium.

Exploratory Data Analysis (EDA):
Performed EDA on IMDb 50K movie reviews dataset released by Stanford. Analysed distribution of target variable (sentiments), distribution of reviews over ratings 1-10, Frequency of reviews length distribution for positive and negative reviews, Frequency distribution of reviews wordcount, word cloud for negative & positive review groups.

Data (Reviews) Cleaning :
Identified Reviews as text column and performed NLP techniques to clean the review data.
Unnecessary information like HTML links, non english characters, english stopwords were filtered out.
Further a list of frequent words in both positive and negative reviews were filtered out as they will not help in classification of the sentiments.
Also, lemmatization was performed to transform variant forms of the same word into a single word.

Sentiment Analysis :
Transformed text into BOW and TFIDF vector representations.
Sentiment Analysis is classification task as we predict two categories - positive or negative.
Modeled three machine learning classifier models - Naive Bayes, Decision Tree and Random Forest to predict the sentiment.
Also, developed Deep Learning models like LSTM and RNN to compare its performance with them.
Performed comparative analysis to find the best model performance by evaluating every model on classification metrics.
Random Forest with TFIDF vectorization had best overall performance, not just in accuracy but also by analyzing confusion matrix and classification report.
The final model was trained on the entire data and used as Sentiment Analyzer for unseen data.

Topic Modelling :
1. Latent Dirichlet Allocation (LDA) with BOW and TFIDF vector representations.
2. Non-Negative Matrix Factorization (NMF) with BOW and TFIDF vector representations.

Tech Stack :
python, pandas, numpy, seaborn, mathplotlib, scikit-learn, gensim, nltk, pytorch.

NYC TLC DATA ENGINEERING

This project focus on NYC TLC Data Analysis using Python, GCP Storage, Compute Engine, Mage Data Pipeline Tool, BigQuery, and Looker Studio. Aims to extract insights from the dataset for informed decisions and deeper operational understanding.

Data :
Data is provided by NYC Taxi and Limousine Commission (TLC) on their portal.

Data Engineering :
1. Modeled data using star schema.
2. Performed ETL (Extract-Transform-Load) on the data with the help of Mage Pipeline.
3. Used BigQuery to query the loaded data.
4. Visualized various trends in the data using Looker Studio.

Tech Stack :
python, pandas, numpy, seaborn, mathplotlib, Google Cloud : GCP Storage, Compute Engine, BigQuery, Mage Data Pipeline, Google Looker Studio.

Real-Time Human Eye Gaze Tracking

The project focused on developing a real-time eye gaze tracking system using the OpenCV computer vision library. The primary goal was to accurately and efficiently determine the direction of a person's gaze by detecting their face and eyes in a live video stream. The project employed several techniques and classifiers to perform face and eye detection, followed by image processing to interpret the gaze direction.

Gaze Detection Pipeline :


Machine Learning :
The classifiers named HaarCascade, Dlib, and Caffe were evaluated and tested to determine their accuracy, speed, and robustness in real-time applications.

Deployment :
The trained models were tested using OpenCV application on real-time visual feed.

Tech Stack :
python, OpenCV, pandas, numpy, pytorch.

Multi Class Image Classification of Dog Breeds

In this project, we harnessed the power of computer vision and machine learning to address the intricate challenge of dog breed identification from images. We developed a spectrum of classification models and neural networks, conducting a robust statistical analysis to evaluate their performance. Through strategic methods such as scaling, normalization, bounding box annotation, and augmentation, we not only boosted accuracy but also established a proficient framework for distinguishing and classifying dog breeds with exceptional precision and efficiency.

Dataset :
The Stanford Dogs dataset contains 20,580 images of 120 breeds of dogs from around the world. This dataset has been built using images and annotation from ImageNet for the task of fine-grained image categorization. It addresses a challenging problem as certain dog breeds have near identical features or differ in colour and age.

Machine Learning :
OvR Logistic Regression, Multinomial Logistic Regression, Support Vector Machine, 3 CNN models - AlexNet, VGG-19, and GoogLeNet.

Tech Stack :
pytorch, python, pandas, numpy, scikit-learn.

Image Caption Generation

Image captioning, a pivotal intersection of Computer Vision and Natural Language Processing (NLP), involves the transformation of visual content into concise textual descriptions. This research undertaking aimed to assess the effectiveness of three distinct architectural approaches in addressing this intricate challenge: Encoder-Decoder / Inject Architecture, Encoder-Decoder with Attention Architecture, and Merge / Multi-Modal Architecture.

Dataset :
Flickr8k dataset consists of 8,096 images. These images cover a wide range of subjects, scenes, and objects, making it a diverse collection for research purposes.

Machine Learning :
The Image Captioning System is made up of three main parts: the Image Encoder, the Sequence Decoder, and the Sentence Generator. These 3 components are connected in obtain different architectures that are explored in this project - Encoder-Decoder / Inject Architecture, Encoder-Decoder with Attention Architecture, and Merge / Multi-Modal Architecture. BLEU (Bilingual Evaluation Understudy) are used to measure the quality of generated captions. BLEU compares the generated captions to reference captions to assess how well they align with human-generated descriptions.

Tech Stack :
tensorflow, keras, python, pandas, numpy.

Virtual Stock

Stock Trading Simulation Desktop Application

The "Virtual Stock" project represents a Java Swing Model-View-Controller (MVC) application designed to simulate stock trading with a keen focus on best practices and optimized system architecture. An application that helps users who are new to investing to learn about how money could grow, in the style of "virtual gambling". Similar to virtual gambling, our application will not use real money and investment. Instead it will allow the user to create investment portfolios, try out buying and selling of stock, and various investment strategies to see how they can grow (or shrink) their money with time.

Architecture :
The system follows a robust Model-View-Controller (MVC) architecture, ensuring a clear separation of components and promoting maintainability. The Model encapsulates the application's data and logic, the View handles the user interface, and the Controller manages user interactions, creating a highly modular and extensible system.

Web API Integration :
The application is integrated with a web API to provide real-time stock data, enhancing the simulation's authenticity. This integration not only enriches the user experience but also allows for efficient data storage and retrieval, ensuring that users have access to the latest market information.

Design Patterns :
The project incorporates various design patterns to optimize its architecture. Patterns like Singleton, Abstract Factory, and Factory Method have been utilized to ensure that each component is highly maintainable and extensible.

Tech Stack :
Java, API, MVC, Java Swing, Git.

Interior Illusion

Augmented Reality Mobile Application

"Interior Illusions" is an innovative application that leverages Augmented Reality (AR) to transform interior design and furniture shopping experiences. With a focus on enhancing user perception, this mobile application enables users to overlay 3D objects, including furniture and interior design products, onto their real environment. This AR technology empowers users to visualize and evaluate how selected objects will integrate into their living or working space.

Mobile Application :
The core of "Interior Illusions" is a mobile application developed with Unity. It offers an intuitive and user-friendly interface for users to interact with virtual furniture and interior products from the database.

Admin Website :
The project includes a web-based administration interface developed with HTML, CSS, and Bootstrap, offering admin users the tools they need to manage and update the database and product listings. The server's functionality is enhanced through PHP scripts.

Tech Stack :
C#, Unity, php, Javascript, HTML, css, bootstrap, REST API, SQL, Tomcat, Git, Blender.

Food Burst : Restuarant Management System

Java Swing and Mobile Application

The Restaurant Management System is a comprehensive software solution that seamlessly integrates a Java-based desktop application with the Food Burst mobile app to enhance and streamline restaurant operations. The project followed a well-structured approach throughout the Software Development Life Cycle (SDLC), encompassing Planning, Requirement Gathering, Analysis, Design, Development, and Testing phases. The project enables registered users to access and utilize various online services offered by the restaurant. This app caters to both customers and the restaurant staff, enhancing the overall dining experience and operational efficiency.

Modules :
The project involved the creation of multiple modules, each serving a distinct purpose, to enrich the online services offered to end users. These modules included:
Menu : Allowing users to browse the restaurant's menu and select their desired dishes.
Food Ordering: Facilitating the easy and convenient ordering of food items.
Payment: Streamlining the payment process for user convenience.
Order Tracking: Enabling users to monitor the status and progress of their orders.
Review: Allowing customers to share feedback and reviews of their dining experience.
Table Reservation: Enabling users to reserve tables in advance.
Party Hall Reservation: Offering the facility to book party halls for special events and gatherings.

Java Swing Application :
In the first phase, Desktop Application was developed. Used Java Swing framework to create versatile, cross-platform desktop software. With rich GUI components and event-driven programming, it delivers interactive and visually appealing user interfaces.

Android Application :
In the next phase, an Android app for the above app was developed. The Android application for Food Burst serves as a comprehensive and intuitive mobile tool for enhancing the restaurant's interaction with its customers.

Tech Stack :
Java, Android, HTML, MYSQL.