Contributed to a 28% increase in sales for a major automobile client by redesigning their webpage with introducing personalized services for their online customers.

The end application was fully developed and deployed on Google Cloud platform and AWS. The batch data from the website was feed into Amazon S3 and it was then transferred into Google Cloud Storage for further processing. Mainly, data processing was done in two phases: Batch Processing and Real-Time Processing. The batch pipeline was created to train the model on historical data and Real-Time Pipeline was created for analyzing the online customers' data and predicting the most preferred car models for customers in near real time. The full application was developed within a container using Docker. 

The application was based on two models: Order Prediction Model and Resource Optimisation model.  

​

Order Prediction Model:

Orders data (by day and slot) was collected for the last many years from the client. In addition, few other macro-economic attributes (like weather/events near the store location) were collected which would have influenced intake orders in past. A time-series analysis was performed to get future orders' count.

Resource Optimisation Model:

​

The objective function of this optimization problem was to maximize the online intake orders and the constraints were defined considering limited resources including labor hours, storage capacity in orders provisioning area. Linear programming model was used as a solution for this problem statement.

All different sports and Apparel brands used to perform market analysis before launching any new product into the market or to understand existing products' performances. Sometimes they also want to understand current users' reviews on the products already available in market. In addition, they need to decide on online advertising budget in different channels including Facebook pages, Youtube, Other websites.  

The E-commerce client wanted to know more about common concerns faced by their customers on regular-basis. They provided big dataset of all customers' reviews for last 1 year. In addition, overall customers' sentiments needed to be analysed.

The analysis report was based on two models: Topic Model and Sentiment Analysis model.  

​

Topic Model:

Latent Dirichlet Allocation or LDA model is a kind of Generative Probabilistic Model consisting of a set of composites (documents) made up of parts (words). It was used to understand different abstract topics most frequently discussed by their online customers on the website.

Sentiment Analysis Model:

Customer sentiments could be positive, negative or neutral. We tried to classify all individual comments into either of these three categories. As part of the analysis different supervised learning algorithms were evaluated and the model was chosen based on accuracy and reliability.