Mar 2016 – Apr 2016 · 1 mo · IIT Kanpur

• The project deals with the deployment of any server code provided by a user using a set of configurations. The set of configurations includes various types of architecture for MongoDB database (Simple, Shard & Replica) and on how many servers does the user want the load balancer to work.
• 1. User can easily add more servers during run-time to the load balancer of an already deployed code . Termination of any instance and launching a new instance can be done instantly using UI.
• 2. Docker containers wrap up a piece of software in a complete file system that contains everything it needs to run: code,run time, system tools, system libraries anything you can install on a server.
• 3. Containers isolate applications from each other
• and the underlying infrastructure while providing an added layer of protection for the application.

Jan 2016 – Apr 2016 · 3 mos · IIT Kanpur

• The aim of the project was Object Detection and Classification in surveillance videos using machine learning and image processing techniques. The objects in the videos were auto, rickshaw, car, bicycle, pedestrian, nameplate and motorcycle.
• 1. Data pre-processing was done manually along with image resizing and gray scaling to remove some erroneous frames.
• 2. Object detection using background and foreground separation was done using methods like Frame differencing etc.
• 3. These objects were extracted from each video so as to train the classification models for multiclass classification.
• 4. Features were generated using Histogram of Oriented Gradient (HOG) algorithm, which were the input as the training data for numerous ML algorithms such as SVM, Decision tree classifiers, Random Forest Classifier, Adaboost and KNN

Aug 2015 – Nov 2015 · 3 mos · IIT Kanpur

• The aim of the project was to explore the field of adapting skyline computations on a mapreduce framework.
• 1. A hadoop cluster using a master-slave architecture with 3 physical slave nodes was set up to implement the various algorithms and evaluate their performances along with benchmarking the performances of these algorithms on our Hadoop distributed file system
• 2. The map-reduce paradigm proposed by Dean and Ghemawat in 2008 forms the crux of the algorithms that we have used.
• 3. Algorithms like BNL, SFS and Bitmap were adapted to the mapreduce framework. Also, partitioning using angular partitions instead of grids was used to ensure each mapper instance getting some skyline nodes

May 2015 – Jul 2015 · 2 mos · Bengaluru, Karnataka, India

• Worked as an Intern in Seller Registration and Engagement Team at Amazon,Bangalore.
• 1. The Catalog Builder provides Seller with UI for getting his product data exported into excel file as per amazon style guidelines
• 2. Seller will be able to create a complete feed with variations. Seller can also upload a raw file and get it converted in amazon format and can manually enter data,validate the data and see the errors in the UI.
• 3. Seller can correct the data and regenerate the file, can also see the instructions and examples of all the attributes and can delete or duplicate some data.

Aug 2014 – Nov 2014 · 3 mos · IIT Kanpur

• The aim of the project was to add an extra dimension to the traditional game of Tetris and implement it using C++ , OpenGL and OpenAL.
• 1. Designed a Heads-up Display (HUD) - the HUD will display the current player score, game title and a side window displaying the next block.
• 2. Implemented texture mapping for the blocks.
• 3. Used ambient lights and particle effects to add visual appeal and added sound effects to game.
• https://github.com/luckysahani/Tetris3D-OpenGL-Game

May 2014 – Nov 2014 · 6 mos · IIT Kanpur

• Implemented Image Processing on Video Surveillance of IIT
• Kanpur . The aim of the project was to detect the number plates on every moving
• object and categorize them . The unwanted frames were removed by separating foreground from background using Method of Gaussians in OpenCV (Implemented the
• algorithm for an improved adaptive background mixture model for real-time tracking
• with shadow detection).
• 1. Studied various methods for background subtraction and motion detection (optical flow) for selecting candidate frames that contain useful data in the surveillance video.
• 2. Implemented a real-time system for adaptive background subtraction using Gaussian Mixture Model
• 3. Implemented Viola-Jones object detection framework for vehicle classification using OpenCV
• 4. Extracted candidate license plate areas from the images and enhanced those using morphological operations as a first step towards Optical Character Recognition.
• 5. Improved the background subtraction to work better in large illumination and weather changes

May 2014 – Jul 2014 · 2 mos · IIT Kanpur, Kanpur, India

• 1. Worked at a start up jutja.com, a project management website with mind-mapping tasks.
• 2. Designed the user interface of the website and was the front-end developer of the website.
• 3. Implemented the mind-mapped visualization of tasks using Vivagraph library in JavaScript.
• 4. The website uses single-page interface to avoid reloading of webpage.

May 2013 – Jul 2013 · 2 mos · IIT Kanpur

• Studied and Implemented various Supervised Machine Learning Algorithms like SVM, Forest Tree, Naive Bayes etc. to classify Emails into pre-defined categories based on their content.
• 1. The project involved creating a mail classier based on bag of words model using publicly available Enron mail database.
• 2. Aimed at building an automatic categorising routine using the classier.
• 3. Studied and implemented various supervised learning algorithms like Multiclass-SVM, Naive Bayes, Random forests. Also used basic natural language processing techniques like POS-tagging, stemming and separating word clusters.
• 4. Achieved an accuracy of upto 80% on balanced datasets but due to high variance observed in such tasks accuracy down to 45% was also observed on some datasets.
• 5. The main testing and classication module was built using python, python-nltk, scikit-learn, weka. A prototype for Gmail was also built using Google App script.