NIKHIL NAN

Designed and built a deep learning model that generates spoken image captions to aid visually impaired individuals. Based on the Flickr8K dataset, the system leverages a CNN-RNN architecture for caption generation and integrates an attention mechanism for enhanced contextual relevance. The generated captions are then converted to speech using a text-to-speech library, making images accessible through auditory descriptions. This innovation combines computer vision, natural language processing… Show more

Designed and built a deep learning model that generates spoken image captions to aid visually impaired individuals. Based on the Flickr8K dataset, the system leverages a CNN-RNN architecture for caption generation and integrates an attention mechanism for enhanced contextual relevance. The generated captions are then converted to speech using a text-to-speech library, making images accessible through auditory descriptions. This innovation combines computer vision, natural language processing, and accessibility, extending the capabilities of the "Show, Attend and Tell" research paper. Show less

Implemented a sophisticated gesture recognition system designed to integrate with smart TVs, offering a hands-free user experience. Utilized Conv3D and Conv2D+LSTM/GRU deep learning models, trained on hundreds of videos capturing various hand gestures. Enabled functionalities such as channel changing, volume control, pausing, and fast-forwarding, all through hand gestures, as detected by the smart TV's webcam. This innovation not only improves user engagement but also sets a new standard for… Show more

Implemented a sophisticated gesture recognition system designed to integrate with smart TVs, offering a hands-free user experience. Utilized Conv3D and Conv2D+LSTM/GRU deep learning models, trained on hundreds of videos capturing various hand gestures. Enabled functionalities such as channel changing, volume control, pausing, and fast-forwarding, all through hand gestures, as detected by the smart TV's webcam. This innovation not only improves user engagement but also sets a new standard for smart device interaction. Show less

Developed an AI-based classification system to expedite and enhance the accuracy of melanoma skin cancer diagnosis. Utilized advanced image processing techniques on dermatoscopic images to train a Convolutional Neural Network (CNN) for identifying nine types of skin cancer. The project aimed to significantly reduce the time lag— from a week to just a couple of days—involved in the traditional biopsy and examination process. Achieved an increased prediction accuracy, supplementing the existing… Show more

Developed an AI-based classification system to expedite and enhance the accuracy of melanoma skin cancer diagnosis. Utilized advanced image processing techniques on dermatoscopic images to train a Convolutional Neural Network (CNN) for identifying nine types of skin cancer. The project aimed to significantly reduce the time lag— from a week to just a couple of days—involved in the traditional biopsy and examination process. Achieved an increased prediction accuracy, supplementing the existing 75-84% diagnostic accuracy rate, thereby offering a transformative impact on early-stage detection and treatment for millions of individuals. Show less

Executed a project to improve the ability to detect fraudulent transactions in a credit card company. Focused on a challenging, imbalanced dataset of European cardholder transactions from September 2013, applying a variety of machine learning algorithms such as Naive Bayes, KNN, SVM, Random Forest, GBM, XGBoost, and LightGBM. Successfully achieved an AUCPR score of 0.86 using the XGBoost model, exceeding the industry benchmark of 0.85 and increasing the accuracy of identifying fraudulent… Show more

Executed a project to improve the ability to detect fraudulent transactions in a credit card company. Focused on a challenging, imbalanced dataset of European cardholder transactions from September 2013, applying a variety of machine learning algorithms such as Naive Bayes, KNN, SVM, Random Forest, GBM, XGBoost, and LightGBM. Successfully achieved an AUCPR score of 0.86 using the XGBoost model, exceeding the industry benchmark of 0.85 and increasing the accuracy of identifying fraudulent activities. Show less

Utilized historical customer data encompassing demographics, services availed, and monthly expenses to build a predictive model for customer churn. Developed a binary classification algorithm that accurately identifies the likelihood of a customer switching to a different service provider. This AI-driven approach has empowered the business to proactively address customer retention, thereby reducing churn rates and enhancing revenue stability.

To develop a high frequency voltage injection system to inject high frequency voltage signals with low magnitude to a high voltage system having partial discharge and to investigate the effect of the injected signals on the behaviour of partial discharge.