Machine Learning | Shikhar Agrawal

ZEVA

Sun, 01 Nov 2020 09:52:15 +0000

Zeva is Zensar’s voice-based enterprise personal assistant. With features of enterprise integration, email integration, and Google search integration. Zeva has an in-built self-learning capability based on its usage pattern. Built on an AI-based Q&A system, Zeva maintains context in the conversation using voice authentication and offline speech features.

Feedback based continuous learning: designed and implemented a feedback loop for improving the answers by ZEVA. Feedback is provided through a like/dislike button. Trained an ML model to continuously learn from feedback.

Unstructured text retrieval: implemented a custom fine-tuned model based on BERT to retrieve answers from unstructured text. Improved accuracy by approximately 48% compared to the previous IR model.

Refactoring monolith to microservices: collaborated with the team to convert the previous monolith version of ZEVA to the current microservices architecture.

Social Distancing Tracker

Sun, 01 Nov 2020 09:51:04 +0000

Implemented a social distancing tracker using Computer Vision (CV) to track the violation of social distancing norms. It requires a one-time calibration step. IDs of people violating the norms get logged in the SQL database. This solution is integrated into various retail and manufacturing clients.

Image Similarity Solution

Sun, 01 Nov 2020 09:43:52 +0000

Implemented an image similarity solution for a major UK based retail client. Multiple algorithms were built, tested, and compared based on accuracy, response time, and match rate for various image transformations.