PDF copies of the publications can be downloaded from an article’s page. If you have any questions, shoot an email.
Our HP Inc. forecasting framework was featured in Foresight: The International Journal of Applied Forecasting (Issue 79, 2025) and recognized as a finalist at the International Institute of Forecasting’s Foresight Conference. 🔗 PDF
HP Inc. replaced manual and statistical forecasting with a machine learning (LightGBM) model to improve demand prediction accuracy across 18,000+ print products. The model has been deployed enterprise-wide, with demonstrated business value and principles for scaling ML in large organizations. 🔗 PDF
My Ph.D. dissertation (University of Tennessee, 2025) develops a machine-learning-driven demand forecasting framework implemented at HP Inc., improving forecast accuracy by 34% and reducing inventory by 28%. 🔗 PDF
We proposed a novel end-to-end approach, the Neural Lagrangian Selling (NLS) model, to improve Guaranteed Delivery (GD) advertising by concurrently predicting ad impression inventory and optimizing contract allocation. The model incorporates a differentiable Lagrangian layer and a graph convolutional neural network to enable direct optimization of allocation regret and effective handling of various allocation targets and constraints. 🔗 PDF
Invento Robotics by Balaji Viswanathan is probably one of the most famous start-ups in the Indian robotics space. Their flagship robot Mitra was used at a high profile Global Entrepreneurship Summit in October 2017. This marketing case study is on designing their marketing plan. This research work was selected into The Case Center 2021 Competition in the Hot Topics category. Colloquially, this competition is known as the Oscars of case studies.
We applied a dynamic Gaussian process model to predict coverage for novel Malaria vaccines in 78 countries. Using publicly available WHO data on coverage of nine vaccines, we developed localised models for countries grouped using the human development index (HDI). We deployed convolutions of standard GP models with weights determined using singular value decomposition of time-series response matrix. 🔗 PDF
My first academic publication: a peer-reviewed book chapter on statistical modelling using Gaussian processes. We reviewed several GP models and correlation structures, and methods to handle numerical instabilities due to near-singular matrices. Finally, we reviewed several algorithms developed specifically for analysing big data obtained from computer simulators. 🔗 PDF