Vishwam Case Study 1: Highly Optimized TensorFlow Solution
Vishwam develops an extensible emotional AI platform that allows for extraction of hidden and subtle information pertaining to intent, personality, and expression in the communication data streams of text, audio and video files. They develop a hybrid physics-based pseudo-AI network with the key value proposition being minimal amount of data required to create predictable emotional AI models. This self-learning platform can be accessed through Rest-API’s by any application for analyzing text, audio and video feeds in real time, to extract human emotions & personality information.
Vishwam needed a massively scalable solution for its computer vision models. Although, Vishwam already had very efficient and accurate models, the scalability requirements to billions required that the processing pipeline be highly optimized.
Trillo helped Vishwam migrate to a GPU-based solution. In order to minimize impact on the existing implementation, Trillo used TensorFlow Serving over gRPC. It identified the optimal combination of GPU, CPU, memory, cluster size to maximize resource utilization.
As a result, the estimated cost savings were upward of 40%.