Projects Related to Experience & Learnings along the way

Revenue Impact on Games through GamePass Availablity

Associated with Microsoft Xbox

As a Data Science Intern at Xbox, I was involved in Game Pass Impact Revenue Analysis Across Games

• Game Pass Impact Analysis: Analyzed the impact of Xbox Game Pass on game consumption and player reach, identifying potential declines in purchases and unearned recurring revenue.
• Revenue Prediction: Predicted potential loss in base game revenue using CausalML models, focusing on unearned recurring revenue from Game Pass titles.
• Modelling: Trained Time Series forecasting predictive models for understanding treatment effect of game pass availability and also used CausalML models, such as T-learner to understand CounterFacual of Game Releasing in Game Pass.

Fuel Economy Optimization

Associated with Cummins Inc.

Fuel Economy Optimization, recommend trimmable parameter combinations that maximize fuel economy for customer fleets.

• Clustering made clusters of the trips of the unit on param basis, these non-trimmable engine data include duty cycles and driver behaviours. Through this, trips can be categorized into different sectors for better recommendation of trimmable params, as for the new trip, this will help outline the behaviour and take out the similar trips, through Similarity.
• Regression developed accurate boosting models to predict fuel economy using trimmable (engine specs) and non-trimmable (duty cycle, driver behaviour) engine data for each cluster.
• Developed the pipeline, with three main principles, i.e., Clustering, Similarity & Regression, to predict the Optimized Fuel Economy on fleet and unit level, for the recommendation of trimmable parameters, based on prediction made.

Oil Reset Event Detection of an Engine

Associated with Cummins Inc.

Oil Reset Detection, detection of oil change for the Engine specific unit based on the parameters present in the pipeline, which will remove the dependency of customer manual inputs for the reset events, in the production phase.

• Devised a data pre-processing approach, with appropriate filtering of data with the help of the subject matter experts.
• Trained a classification model, with a leverage of vicinity factor for the prediction of oil reset events.
• Implemented a Business Logic based on the customer and service literature, for Minimizing False positives for the model.
• Created a Range Metric for the evaluation of model based on customer input.

Fuel Cell Electrode – Crack Classification

Associated with Cummins Inc.

Initiated collaboration of Hydrogenic Team and AAI team in Cummins. With the objective to detect crack defects in x-ray images of fuel cell electrodes we have experimented with both statistical models and state of the art deep learning models used today for image classification purposes. Applying Non-Destructive Evaluation techniques like X-rays for image acquisition of Hydrogenic Fuel Cell Electrode, we engineered numerical features using the right shift in pixel distribution which helped our statistical model to classify defected and non-defected parts with an accuracy of 88%. For Inferencing, a graphical user interface was developed to collect the batch and visualize the results with report creation functionality, for the ease of model use.

Video Classification into Academic and Entertainment using Subtitles

Research publication Link

Research associated with SRM University

This paper describes our work on building a Classification system for video subtitles. As an example, we report on our evaluation results for two TV genres -Academic and Entertainment. Through this implementation, the user can classify the videos into academic and entertainment in any software or system, which will help the user to differentiate the videos properly into useful and distracting content. This enables applications like social networking and instant messaging apps to filter distracting content, thereby enhancing productivity of users.