Artificial Intelligence Fellow. Insight Data Science

​2010 - present

​

• Built an end-to-end machine learning pipeline for 3D reconstruction of objects from single or multiple 2D views of images integral to AR/VR tasks and computer graphics.

• The two-stage pipeline consisted of semantic segmentation for extraction of objects from images and a 2D encoder – 3D decoder network for generation of 3-dimensional objects.

• Trained the network on the ShapeNet database using CG-generated 3D CAD models and rendered views on AWS.

• Obtained an average IoU of 0.681 for the 11 categories of objects which outperformed the previous state of the art models.

• Productionized the Dockerized model and served it as a web application using Streamlit.

• Federated learning & differential privacy: Developed a python machine learning software to predict the credit card default risk using a federated learning method with PySyft. Measured the effect of privacy on the accuracy of the model predictions with standard ML models.

​

Graduate Researcher at the University of Iowa

​2014 - 2020

​

● Synthetic Microstructure Generation using Transfer Learning:

1. Implemented a computer vision software in TensorFlow for the reconstruction of realistic material microstructures via transfer learning method applicable to various types of microstructures.

2. Obtained a less than 5% reconstruction error calculated using the two-point correlation method and probability density distribution of microstructural descriptors.

● Machine Learning-based multiscale modeling:

1. Used a Gaussian Process Regression method to learn the chemical response of shocked explosives from micro/nanoscale computational simulations to build surrogate models.

2. The key achievement of the work is that it used small data machine learning techniques to couple the physics of the material to the microstructural defects.

● Learning from Simulations:

1. Designed a CNN architecture to predict the physio-chemical behavior of explosives as time-series data and obtained a prediction accuracy of 75 %.

● Material Microstructure Quantification:

1. Created a codebase for statistical quantification of material microstructure from Scanning Electron Microscope
   images using a level set-based technique, methods were published in a peer-reviewed journal.

2. Developed a novel technique to link material microstructure to the physio-chemical response of shocked multiphase flows.

● Synthetic Materials by Design:

1. Developed a GAN based technique for the generation of controllable microstructures and performed computations on the synthetic microstructures to obtain optimized desired material response (on solid explosives).

​

Senior Mechanical Engineer at Hindustan Construction Co.

​2011 - 2013

​

Planned, managed, and designed mechanical systems such as large-scale industrial equipment, batching plants, crushing plants, and turbine systems for the construction of Teesta Hydro Power Project with a capacity of 160 MW (4X40 MW) in a record time of 196 days using the roller compacted concrete technique (third time in India).