Surya Shravan Kumar Sajja

I am a  Senior Research Scientist in the AI for supply chain group at IBM Research Lab, India. I was a research scientist in the Optimization, Control and Decision Sciences group at IBM Research Ireland from 2013 to 2016.  I received my doctorate in mathematics from Hamilton Institute, Ireland and my masters from Indian Institute of Technology Bombay, Mumbai, India. I was also a postdoctoral researcher in University of Notre Dame and a visiting researcher in TU Berlin, EPFL and ETHZ. My research interests include explainable and robust AI, choice theory, causality, demand forecasting, stability theory, stochastic optimization and interconnected systems. I published my research in journals like Automatica, IEEE Transactions on automatic control, Systems and Control, Transportation science and in conferences like AAMAS, ACC, CDC, EDGE, ADT.

Select Industry Experiences

• Explainability for optimization. This work has been initiated to address the explainability needs of IBM’s AI apps. Here, we are primarily focused on providing explainability to IBM’s MAXIMO Graphical Scheduler for Large Projects (GSLP) and Sterling’s Omnichannel Fulfillment Optimizer (SFO).

1. For GSLP, We developed the resource constrained critical path evaluation module as an explanation tool for the schedules generated by ILOG constraint programming solvers used by MAXIMO-GSLP. This work has been extended to align with large scale maintenance schedules from clients like Airbus. Currently, we are focused on generating meaningful visualizations for such large-scale critical sequences and What-If analysis of CP Optimisers used for scheduling

2. For SFO, we developed the alternate solution-based explainers to explain the optimal solutions generated by CPLEX engine used by SFO. We also developed expert guided alternate solution module, which has been further used to provide What-If plots to SFO users. Currently, we are working towards improving our solutions using more realistic synthetic data and visualizing alternate solutions more effectively for lay users.

• Explainable new product forecasting and AI based interventions for pre-season decision making in fashion retail:  Fashion industry is heavily influenced by seasonal market trends and 70 to 80% products introduced every season are new in nature. New product development and design in fashion industry includes stakeholders like designers, buyers, merchandizers and planners. Our goal was to build an explainable new product forecasting tool with capabilities of interventional analysis such that all the stakeholders (sometimes with competing goals) can participate in collaborative decision making process of new product design, development and launch. 

• Stock Allocation and Market Demand Prediction based on Demography: Was part of team that developed methods to extract market sentiment using text (reviews) and image data. I developed explainable models to forecast market sentiment of SKUs at different hierarchies for SKUs and locations. These spatio-temporal models were further used for pre-season stock allocation and mid season stock transfer.

• Price Prediction using Collaborative Cognition: Developed multiple multivariate statistical models to predict prices of agricultural commodities in different markets of south India. These models were further enhanced using a Collaborative AI framework which iteratively improves prediction by combining predictions from multiple domain experts and multiple models.

• HVAC Optimization and Occupancy Measurement for Green Buildings: Optimization of Heating and Ventilation and Air Conditioning systems needs accurate measurement of occupancy. We developed privacy preserving Wi-Fi based localization system at an enterprise scale and it was used to generate accurate occupancy measurements. These measurements were used to generate optimal sequence of temperature and pressure set points for HVAC components.

• Vulnerability prediction models for elderly care: Vulnerability assessments, are critical in outcome-based personalised care management.  I developed predictive models (based on a combination of a Markov Model and a Bayesian Network) to predict Vulnerability of individuals and asses their needs. This model was built by combining opinions of multiple experts and multiple datasets which were incomplete and partially overlapping.

• Personalized routing based on driver behavior modeling: Developed driver behaviour models, using data generated by sensors and cameras mounted on the cars and on board diagnostics (OBD) data of cars. These models indicated driver’s capabilities on different road types and it was used the generate the best possible route for a given driver with a source and destination in the city road network.