- Designed and implemented an iterative clustering algorithm similar to kmeans for item grouping in meta auction marketplace. - Algorithm iteratively minimized the error between optimal cost per click and the realized CPC after applying bidding multiplier constraints. - Algorithm jointly identified the bidding multipliers as well as the cluster assignments.- Mathematically proved that the clustering algorithm has the monotonically non-increasing property of the objective function.- Achieved a 15-20% improvement in RMSE value depending on the point of sale. - Built a pyspark error monitoring library for prediction models used in Brand Expedia meta search bidding algorithms.- A baseline was also implemented using historical averages.- Implemented different error metrics such RMSE MAE and Log Loss to be monitored each day on ~4.5 million data points across 24 points of sale.- Incorporated with the existing Jenkins CI pipeline and set up test cases with 70% code coverage.- Implemented a multiplicative bias correction to different segments of the gross profit per click (GPC)predictions.- This allowed the marketing campaigns to switch from 30 day attribution to same session attribution,without losing market share.- Won A/B tests with 7-10% GP lift at the same return on advertising spend.- Provided other data science and software solutions to the campaign managers to drive Brand Expedia metasearch campaigns that has a spend of more than 300k USD per day.

- An undirected probabilistic graphical model was developed to infer hidden states in a network.- Neighbor influence and temporal dynamics were modeled through exponential factors.- Variational inference scheme was proposed based on the mean field approximation.- A supervised training method was developed using the stochastic gradient based maximum likelihood approach.- Implemented a Gibbs sampler from the joint distribution for obtaining stochastic gradient.- Achieved better classification accuracy compared to Support Vector Machines (SVM) and Hidden Markov Models (HMM) with ROC-AUC greater than 0.9- A model was developed to predict user activities in a social network by observing only a subset of users.- Observed user activities are modeled as counts from a specific action category.- Developed a novel unsupervised training method based on variational expectation-maximization and stochastic gradient.- For a Twitter dataset containing health sentiments, lower Akaike Information Criterion values were obtained compared to Poisson hidden Markov models and mutually exciting point process.- 34.9% reduction in mean absolute percentage error was obtained for the same dataset for predicting user activities observing only 50% of the users.- A greedy sensor selection algorithm was proposed for estimating probabilistic graph signals.- This is equivalent to variables selection in a Gaussian Markov Random Field for estimating unobserved random variables. - Repeated matrix inversions were made efficient using Sherman-Morrison formula.- Improved the time complexity of an existing algorithm from O(n^4) to O(n^3).- Experiments demonstrated that the proposed method runs in 10 % of the time of the next best method.- Proposed method gave the best minimum mean squared error performance among existing methods.- The method was further extended for time varying graphs assuming a single edge is modied at each time.

Carried out research in the field of evolutionary design of mechatronic systems using bond graph and genetic programming. Develped MATLAB code on extracting state space model of a system represented in bond graph.

Taught the module, Computing for Engineers for the first batch of students in Electronic Engineering. Developed lab sheets and lecture notes to make the students familiar with MATLAB software.

Gained good understanding on avionics, navigational aids and Radar (PSR & SSR). Completed a project to send Radar data through internet on a secure socket layer.