Shashank Sharma

Experience

• 

  Tata Motors

  May 2012 - Jun 2012

  Summer intern

  the name of the project was ‘localisation of lcv splits’, done under the guidance of the department of vendor development aimed at cost reduction via development of the parts from local vendors. the project was a success and it decreased the costs, risks and process delay involved in long distance and inefficient transportation of parts.

• 

  INDIAN RAILWAYS

  Jun 2012 - Jul 2012

  Summer intern

  one month long internship at railway research, development and standardisation organisation (rdso), lucknow. the project involved the ‘finite element analysis of the transverse screw coupling’. the aim of this project was the overall analysis of the specified part under the working conditions described by the modified wdg4 locomotive.

• 

  YML

  Jul 2013 - Jun 2015

  Senior software engineer

  - design and build advanced applications for the ios platform, collaborate with other teams to build new features, continuously discover, evaluate, and implement new technologies to maximize development efficiency- created a set of libraries to aid fast and efficient development and remove some unreliable and inefficient third-party dependencies- created several successful pocs with very short deadlines- part of several projects, all done from scratch, of which two projects reached top of the app store in their categories show less

• 

  Srvs, Inc.

  Jul 2015 - Feb 2016

  Senior software architect

  srvs is a blue collar job portal that allows users to view jobs, apply to them, manage resumes, schedule interviews, message for convenient connection of users and employers. on the employer’s side the application allows them to post jost, receive, shortlist, schedule interview, hire applicants and manage employees.- founding software engineer and lead ios developer- coordinated work between teams spread across timezones- developed the complete ios application from scratch- assisted in product development via materialising various concepts into features show less

• 

  Spocket

  Feb 2016 - Aug 2016

  Co founder & technical lead

  spocket (spocketlive.com) is a social platform for sports enthusiasts that features crowd curated content and predictions on match events via virtual currency- lead a agile team of developers to create this platform based on user feedback- designed and created the android application from scratch and supervised the development of backend and website- aided in digital marketing of the product over social media platforms

• 

  Self employed

  Aug 2016 - Nov 2016

  Deep learning enthusiast

  decided to take some time off and dive into deep learning- read a lot of research content to stay updated with the latest developments and created a solid foundation for myself in linear algebra, probability theory, information theory, and deep learning- work with advanced frameworks like numpy, tensorflow, keras, scikit-learn- produce scaled down versions from research projects as toy samples implemented on data sourced from kaggle, academictorrents, etc- image classification; image generation, text generation; interested in unsupervised learning show less

• 

  Indian Institute of Technology, Kanpur

  Dec 2016 - Oct 2017

  Researcher

  ▪ involved in research of application of unsupervised deep learning methods to the field of ​computer vision ​(images & videos) or other spatio-temporal data through generative adversarial networks.▪ the research was focused on development of methods to effectively capture all the modes with a theoretical guarantee. the resulting paper, ​‘no modes left behind: capturing the data distribution effectively using gans’ was accepted for an oral presentation at the proceedings in ​aaai’18​. sample results and additional links can be found at: ​https://github.com/shashank879/logan▪ the results were evaluated on several toy and natural datasets; celeba and another dataset prepared by extracting frames from a surveillance video; the latter heavily encouraged mode collapse. it was shown that the proposed method was indeed superior and tackled the problem of mode missing/collapse directly.▪ earned a ​google grant​ as the lead author of the paper.▪ the models were written in python using tensorflow, scipy, numpy, etc that deployed on distributed systems. show less

• 

  Active.Ai - A Gupshup Inc Company

  Nov 2017 - Dec 2019

  Researcher

  ▪ implementation of deep learning methods to develop an ai stack for natural conversations forfinancial services.▪ high-level tasks involve ​text reformulation using deep reinforcement learning,​ ​abstractivesummarisation​, ​machine comprehension​, ​natural language generation & augmentation​, natural language understanding​, ​context handling for multi-sentence conversation, translation/paraphrasing​,​ rl based reformulations​, etc.▪ achieved a ​world ranking of 14th and 10th on squad v2.0 and v1.1 datasets respectively at the task of machine comprehension.▪ low-level tasks involve ​sequence classification​, ​sequence tagging in the context of tasks like semantic role labelling, named entity recognition, etc. using ​lstm​ and ​cnn​ based encodings.▪ the modules have been written as ensembles of deep learning components with modern mechanisms like ​attention​, ​self-attention​, ​differentiable neural computers​, ​pointer-generator, reinforcement learning assisted optimization​; and inspiration from state of the arts to perform specifically for our domain. show less

• 

  Google

  Jan 2020 - Jan 2022

  Pre doctoral researcher
• 

  University of Bath

  Oct 2021 - now

  Phd student