Question Answering Tasks Using BERT models

• Experimented with different BERT based models such as DistilBERT, ELECTRA and RoBERTa on SQuaD 2.0 dataset (Stanford Question and Answering) to find the model that worked best with minimal computational resources.
• Developed a web application that serves the best performing model to answer user-fed questions based on a context.
• Link

Demand Forecasting for Retailers using Weather

• Built a Demand Forecasting Platform for Fashion retailers which included weather as one of the factors
• Used Time Series Analysis techniques such as ARIMA and tools such as FbProphet to understand seasonality, trends and patterns in the data
• Link

Fake News Detections

• Developed a Bi-directional LSTM model to classify Fake news using a combination of feature vectors obtained from the title and the body of the news article. Trained the model using distributed deep learning on Spark
• Scraped various satirical and real news website using a crawler built on scrapy, cleaned and processed the data using NLTK. Achieved 0.611 as the Precision and 0.69 as the Recall value
• Link

Generalized Zero Shot Object Detection

• Developed a model that detects objects not seen in the training set, using Generative Adversarial Network and Variational Encoders
• The model consists of a Generator, which generates visual features given class embeddings and a regressor that maps each visual feature back to its corresponding class embedding
• Link