Sign Language Recognition

The objective was to build a recognizer that would convert signs from the American Sign Language(ASL) and convert it into text. It takes an image as an input(or captures images from a camera), recognizes a hand in the image and classifies the sign(if there is one) into 26 letters of the English alphabet. This was achieved by leveraging Python and OpenCV to use transfer learning on the YOLOv3 Object Detection model. The model achieved a mAP(mean average precision) of 60.6.

Tech Stack - Python, OpenCV, Object Detection, Transfer learning, Convolutional Neural Networks

Face Recognition

The objective was to encode known faces, save them, and have the recognizer use these saved encodings to recognize people from images or video.

Tech Stack - Python, OpenCV, Face Recognition, Convolutional Neural Networks

ML based Item Weight Prediciton - Walmart Global Tech

Today in this highly digitized world, data accuracy is paramount, especially in the supply chain world where even a small percentage of data discrepancy could hurt business. In order to leverage artificial intelligence to provide solutions to remove item meta-data discrepancy in the product catalogue that is fed into the Sam’s Club sourcing engine (FASTr), a Machine Learning model was developed for predicting the weight of an item in the product catalogue.
The dataset consisted of 200,000 items whose metadata was used for training and testing the machine learning model. The main features used were the item's name, category, manufacturing info specification, short description, brand and keywordsfor predicting its weight. Several preprocessing methods, Vectorization techniques and Machine learning and deep learning algorithms were tested. The best model had an R2 score of 0.93.
This model will be integrated as a service into the sourcing engine pipeline to cross-check the item weight meta-data in the database, and flag discrepancies.

Tech Stack - Python, PySpark, Azure Databricks

Sentiment Analysis for Product Reviews

For this project, text reviews were processed using machine learning algorithms to classify them as positive, negative or neutral (Sentiment Analysis). The research compared multiple Rules based approaches(such as TextBlob, VADER) as well as Feature Extraction techniques(Logistic Regression, Support Vector Machines) based on accuracy and the macro F1 score, in order to determine the best approach to perform Sentiment Analysis.

Tech Stack - Python, Machine Learning

Personal Website

Digit Recognizer

The Android application in Java has a pre-trained tensorflow model which was trained on the MNIST dataset. Any single digit number drawn on the canvas can be recognized by the model with 97% accuracy.