Apr 2020 – Nov 2023 · 3 yrs 7 mos · Delhi, India

• Basil was born from a passion to make clean, unprocessed food accessible to everyone. Recognizing the dilemma of choosing between oily cheap options and expensive diet food while on the go, the founders aimed to be a trusted brand offering affordable, preservative-free daily needs, revolutionizing convenient and healthy eating.

Mar 2018 – Jul 2018 · 4 mos · Boston, Massachusetts, United States

• Meta-learning in Deep Convolutional Neural Networks
• Worked on Neural Architecture Search and Metamodelling in deep computer vision.
• Worked on developing deep neural networks robust to adversarial attacks for image classification.

May 2017 – Jul 2017 · 2 mos · Mumbai, Maharashtra, India

• Attention Models for High-resolution Chest XRay Diagnosis
• Developed attention models inspired by early spatial attention in the human visual system, so as to achieve interpretable results and be able to process high-resolution images without losing important information.
• Combined spatial attention model with object-level attention to model the feedback connections in the human visual system.
• Improved on the baseline methods from 76% to 82% binary classification accuracy for cardiomegaly detection on a private dataset using only spatial attention.

Jun 2016 – Nov 2016 · 5 mos · Paris, Île-de-France, France

• Instance Matching and Object Discovery in Artwork
• Discovered single-view instance matches for small objects (e.g. huts, boats) using Fast-RCNN in an unannotated dataset of artworks from Jan Brueghel's workshop. Studied the effect of network architecture, scaling, bounding box proposal algorithms, and preprocessing on matching quality.
• Applied adversarial learning with synthetic artistic styles to generate style-invariant features for matching artworks across different styles.
• Combined adversarial training with multi-task training in a hybrid CNN architecture to obtain more robust feature representations.
• Achieved state of the art on Artist, Genre, and Style classification on Wikipaintings and Rijksmuseum
• datasets using CNNs trained for content and style multi-task learning.

Dec 2015 – Jan 2016 · 1 mo · New Delhi, Delhi, India

• Created an aerial mapping software for Unmanned Aerial Vehicles using OpenCV in Python, with video mosaicing, deblurring and image super-resolution.
• Implemented superior real time video stabilization by improving Youtube's video stabilization algorithm.