Work Experience

• Developed and released variational models to increase the robustness and realism of agent behavior prediction in simulation.

• Led end-to-end solution, drove experiment design, metric definition, visual analysis, and improved simulated agent realism distributional metrics by XX%.

• Improved trajectory generation and reduced number of collisions by XX% resulting in qualitatively better predicted agent behavior.

• Developed a Hierarchy Aware Multi-label Classification technique with (2.6% AUPRC and 2.85% hierarchical) median percentage gain on six datasets across image/text/audio domains and an 8.87% median percentage robustness gain

• Led a fuzzy match pipeline to detect redundant entities in Google’s JFT, removing 200+ redundancies. Added 50+ important Indian food entities and fixed 100+ incorrect relations. Proposed Indian food image test dataset for Google’s Foodnet

• Demonstrated 15% percentage accuracy improvement on 4-bit quantized ResNet models with checkpoint loading/distillation 

• Developed automatic creation and use of Digital Shelves in winning e-commerce markets, identifying important keywords for mining and ad campaign generation. Used an unsupervised approach to assign keywords to shelves with over 95% accuracy

• Improved relevant keyword volume coverage by 3x and reduced manual set-up time for clients from days to less than 1 hour

• 1. Built Deep Learning models for medical imaging especially for the problem of contextual modeling in Universal Lesion Detection. Processed and penalized each anatomy differently instead of a binary detection problem and outperformed previous state-of-the-art by ~1.5% points sensitivity.

• 2. Built pipeline for explainability in Diabetic Retinopathy classification from retinal images.

• 3. Built workflow for vessel segmentation and Retinopathy of prematurity (ROP) classification from retinal images.

Conducted a critical analysis and survey of Activity Recognition and Video Understanding highlighting not only the key contributions but also the drawbacks of recent approaches. Key insights from the survey were that multimodal cues and better sampling techniques using temporal relations are possible future directions in this domain.

My work was based on creating a part of the pipeline for a hybrid recommendation systems for career planning and recruitment automation.

Created Deep Learning models for player detection, action recognition and inferring other statistics in sports videos. 

Worked on a customized recommendation engine for the company’s API catalogue and created an API for it using flask.

Instructor for multiple introductory as well as advanced courses on Deep Learning for Computer Vision. Taught topics ranging from classical computer vision (HOG/SIFT/Optical flow/BOVW etc.) to deep learning-based approaches to Image Classification, Object Detection, Instance, and Semantic Segmentation, Tracking and Generative modeling