SelfMachines is a modular extensible no-code workflow builder to build, train, manage and serve diverse categories of information retrieval and processing pipelines customized for specific needs of business logics.

Tech Lead in the core perception team. Worked on multi-task learning, early sensor fusion, 2D and 3D object detection and depth estimation.

Tech Lead in the core perception team. Worked on vision pipelines, free-space, LiDAR and camera object detection models, uncertainty estimation.

Designed generative and discriminative deep neural models which scale to large amount of unstructured data available at eBay. Used variants of variational autoencoders to learn compact embeddings for different types of metadata in the areas of ranking and classification.

Did research on Bayesian inference in factor graphs and information theory.

1) Improved the performance of semantic tagging of labels of eBay's titles using deep learning techniques by 20 relative percentage on average for the case of word-vectors as input feature.2) Proposed and implemented a novel method to learn supervised word-vectors for semantic tagging using a deep neural network and available labeled data.3) Implemented a fully configurable neural network library in C++ with a set of features including different cost-functions(sum-of-squares and cross-entropy), different activation functions, different regularization methods(L2 and drop-out), ...

Implemented inference methods with linear complexity in Complex Sparse Graphs in C++.

Teaching Assistant for ECEN303 (Probability Theory)

Found an analytical optimal node density that minimizes the average route BER for a group of wireless AdHoc networks employing minimum distance routing .Studied diffusion processes on dynamic networks and derived a theoretical epidemic threshold for epidemics on switching dynamic networks.