* Working on sensorimotor policy learning for dexterous multi-fingered manipulation using tactile sensors. * Also work on developing and maintaining full stack robotics software.

* Created and Productionized Grasp Prediction Model that predicts both which objects to grasp and how to grasp that object from RGBD camera. * Created and Prototyped IOU Based Tracker for picking faster and for dealing with Dynamic Scene. Above components constitute the significant improvements in speed and reliability of the Singulation Product and the system can perform 2000 PPH with minimal intervention & defects. * Created and maintained regression testing framework for perception pipelines. This pipeline helps catch perception performance regressions early for all product lines at Dexterity. * Prototyped on learning contact interactions for robot maneuvers such as packing / pushing etc. Technical skills: Dex-robotics stack, C++, Pytorch, Robosuite, Open3D.

* Worked on intent and interaction prediction using Graph Neural Networks for Motion Planning team.

* Developed an efficient 2D multi-class detection ( RFCN ) network using camera image data with state of the art detection metrics. Improved the metrics further by using hard example mining. * Developed a novel and highly efficient sensor fusion ( image + lidar) 3D detection model with state of the art results and is a production 3D detector. This work resulted in a patent and was also published in CVPR Workshop 2019. * Extended the similar sensor fusion approach for joint perception and prediction (MultiXNet). This was productionized for joint perception (3D detection) and prediction (short term trajectory prediction). This work is in submission to ICRA 2021. * Mentored 2 interns with one internship leading to a publication in NeurIPS 2019 and another intern getting a return offer.Technical skills: PyTorch, Tensorflow, TensorRT, C++, Propertatory robotics framework.

* Contributed towards #1 KITTI object detection submission at the time (March 2016) based on DetectNet. * Built the image segmentation demo which was showcased by NVIDIA during CES 2016.* Worked on end-to-end driving in simulation (using TORCS) using deep reinforcement learning (A3C). The trained policy exhibited interesting behaviors for nudging other cars for overtaking / efficient turns and recovering from deviations from the road.Technical skills: Caffe, Tensorflow, OpenAI Gym, TORCS.

* Developing consumer review summarization system under the guidance of Prof. Patrick Perry.* Using neural network language models and feature learning for word and sentence representation.Technical skills: Python, Theano, CUDA C/C++.

Worked on user modeling for online / offline customer service and summarizing the web-journey of a customer so that multi-website can be leveraged for predicting consumer intent. During the tenure filed for 2 provisional patents relating to the field. First on distance metric for clustering search terms. Second on extracting customer information from customer interaction chats.Technologies: Hive, Python, R, Mahout, scikits-learn, libSVM.

Worked on text-mining on customer chats, search term segmentation, web-journey visualization and proof of concept projects on user modeling on web.Technologies: Python, R, SQL, Hive, Gephi, NLTK, Stanford NLP Toolkit, scikits-learn, libSVM, Weka.

Worked on smart phone-web services convergence. My work involved design and development of recommendation system based on graph database for personalization of user content.Technologies: Java EE6, XML, REST, NoSQL System (Graph Database), SQL.

Worked on Linux device driver development for RTC and debugging interfaces for Linux-kernel development.Technologies: ARM Core, Linux Kernel, GCC, GDB.