Oct 19 - Jan 21

2D/3D Semantic Segmentation Pipeline
Nomoko AG, Zürich

• This project was developed in collaboration with Dalle Molle Institute for Artificial Intelligence Research (IDSIA) where the task was to perform semantic segmentation on city-scale point clouds reconstructed from our aerial and terrestrial captures with proprietary camera sensors.


• Designed and constructed the machine learning pipelines for ingesting and preparing the data using SfM and SLAM. Training, evaluating, and deploying several Computer Vision models.


• Created pipelines for ingesting unstructured data, such as, image sequences, GPS metadata, other sensor inputs.


• Did a literature review on state-of-the-art classification, detection, and segmentation techniques and deployed an ensemble of methods, including Panoptic-DeepLab, Mask-RCNN, and PointNet.


• Conducted weekly training sessions for the entire software team on several computer science skills, such as Linux commands, Docker, and Kubernetes.

Jun 20 - Aug 20

Database API
Nomoko AG, Zürich

• Developed an API to store semantic annotation data in a model-agnostic format into a Postgres DB using Flask-SQLAlchemy in Python.


• Created the REST API and Click to provide both HTTP endpoints and CLI interfaces to the user.


• It automatically converts segmentation data from various formats, such as polygons or masks, into a standard format and efficiently store it.


• Added the capability of exporting unique datasets by performing a chain of operations, including splitting, merging, filtering, and scaling of several existing datasets, into a model-specific format.


• Finally, deployed MLflow on Linux Servers to train and evaluate several models at scale.

Jun 18 - Sep 18

Scalability of Deep Learning Frameworks
Swiss National Supercomputing Centre, Lugano

• Prepared a report on HPC readiness of Deep Learning Libraries for the CTO of CSCS and recommended candidate libraries to build on Piz Daint.


• Piz Daint, at the time of the project, was the third most powerful supercomputer in the world.


• Deployed a docker container with TensorFlow, Keras, Caffe2, and PyTorch that could utilize the power of 5700 GPU nodes of Piz Daint.


• Ran several weak-scaling and strong-scaling benchmarks on these Deep Learning Frameworks for training Deep CNNs, including AlexNet, ResNet50, ResNet101, ResNet152.

Jan 19 - Sep 19

Real-time object pose estimation using differentiable PnP
Mixed Reality and AI Lab, Microsoft Research, Zürich

• Created an automated pipeline to generate ground-truth data (binary masks and tight 3D bounding-box vertices) from only RGB images of an object, without any use of depth information or Aruco markers.


• The pipeline involves 3D reconstruction, point cloud alignment, background subtraction, and morphological operations for extracting camera intrinsic parameters, sparse 3D reconstruction, tight 3D bounding box, object points, binary segmentation masks, and 6D poses for all the images in the dataset.


• Attempted to improve the accuracy of Microsoft's Single Shot Pose, an approach for simultaneously detecting an object class and its 6D pose, by adding a Perspective-n-Point (PnP) component to the loss function.

Feb 21 - Mar 22

Anti-money laundering and surveillance
Credit Suisse, Zürich

• The project required creating and automating ETL pipelines for cleaning and landing data from multiple sources by writing PySpark transforms in Foundry.


• Aggregated, cross-referenced and combined siloed data from hundreds of disparate data sources.


• Used Entity Resolution to identify overlapping records where joining these data sources was not possible.


• Used Machine Learning to predict risky transactions and mitigate potential money laundering cases.


• Used Slate and Typescript to create dashboards which helped Relationship Manager to make onboarding decisions for new clients.

Sep 18 - Jan 19

AI Driving Olympics Fleet Management Challenge
ETH Zürich

• Designed the Fleet Management Challenge for AI Driving Olympics held at NeurIPS and ICRA conferences.


• This work is a contribution to an open-source project called Duckietown.


• The Duckietown project is a joint effort by several universities including, Université de Montréal, MIT, ETH Zürich, as an effort to drive research in Robotics and AI to help develop various Autonomous Vehicles.


• The Fleet Dispatcher and the Physics Engine for the project are made using ROSPy and the visualization is designed in RViz.

Jan 18 - Jun 18

High Performance Topology Optimization
Università della Svizzera italiana

• Achieved 2x speed in solving the Compliance Minimization problem of Topology Optimization by analyzing the performance of different Linear solvers. Moreover, implemented a Linear Elasticity visualizer using FEniCS computing platform.


• Topology Optimization is a mathematical method that optimizes the layout of material given a set of loads and volume constraints.


• A Topology Optimization algorithm functions by iteratively redistributing volume across the mesh elements such that compliance is minimized.


• The computations in this paper were run on a High-Performance Computing Cluster at the Institute of Computational Science.

Sep 17 - Jan 18

Analysis on Transfer Learning
Università della Svizzera italiana

• Conducted analysis on the accuracy achieved with fine-tuning using Google’s Inception v3 by varying the number of images and classes.


• Performance several data augmentation techniques, including, random cropping, scaling, flipping.


• The datasets used for this analysis were Animals on the Web, Animals with Attributes 2, and NIH Chest X-ray images.

Apr 18

Water Me!
Università della Svizzera italiana

Developed an Arduino-based 3D-printed prototype in 84 hours that senses the temperature, the light exposure, and the soil moisture levels of a plant and makes smart decisions on when and how much to water it.

Jan 18 - Jun 18

Document Clustering
Università della Svizzera italiana

Compared the efficiency of K-means and Agglomerative Hierarchical Clustering algorithms in Document Clustering. Also performed Sentiment Analysis on UCI's Enron emails dataset.

Oct 15 - Oct 16

NBCUniversal Managed Services
InfoCepts, Nagpur

A major part of my role involves providing production support to NBCUniversal for MicroStrategy and Tableau. My team had the responsibility of keeping the servers up and running and the business uninterrupted 24x7 / 365. Additionally, we were responsible for solving any issue or request that our users have.

• Created Tableau, Power BI, and MicroStrategy Dashboards for Ad Sales @ NBCU and Comcast.


• Designed overall Security Architecture for projects and providing Security Roles and Project Accesses to its users


• Raised cases with vendors, including MicroStrategy and Tableau, when we face intermittent issues with their applications. Consequently, resolving these issues in the next version of the software or requesting a patch


• Created Command Manager scripts and procedures for automating admin tasks and saved NBCU 430+ hours of effort annually.

Jul 13 - Mar 14

Master Bank Card
Persistent Systems, Nagpur

Eliminated the hassle of carrying multiple credit cards by developing a software-based process that enabled users to carry just a single bank card and use all their credit/debit cards.

• The project consists of an Android app and a server-side developed using PHP


• The choice of the card that a user wants to use is made through the mobile app


• 2-way authentication is built into the system


• Additional option to remember all your card numbers, pins, CVVs, current balance, and transactions in a secure way

Jul 12 - Nov 12

Maze Solver
Nagpur University

Accelerated the process of testing and debugging Maze-solving algorithms by 120% by creating a simulator using graphics in C. The project was a hit among the local robotics communities and it helped over 300 members in designing their algorithms.

• The project is developed in C programming language, and as a demo it is testing a modified version of Djikstras' Algorithm.


• The program is generalized in such a way that a user can decide the size of the grid, the number of dead nodes, and their positions.


• The software is optimized by making use of Linked-lists and Bitwise Programming for storing the Data.