Revolutionizing Customer Experience: Comcast's AI Platform

Table of Contents

1. Introduction
2. Why Voice is Important
3. The Rise of Voice at Comcast
4. How the Voice Remote Works
5. Challenges in Scaling Data Ingest
6. Tackling the Challenge of Scaling Data Ingest
7. Machine Learning at Scale
8. Pain Points in the Machine Learning Lifecycle
9. Developing and Managing Machine Learning Models
10. The Big Void Between Dev and Production
11. A Unified Analytics and AI Platform
12. Conclusion

Introduction

In today's world, every company wants to build their own ML platform, and Comcast is no exception. In this article, we will discuss the platform for data analytics and ML that Comcast has built to support its mission of reinventing the customer experience with voice and AI. We will explore the importance of voice and how it allows us to fundamentally change the way our customers communicate with our devices and their homes. We will also discuss the challenges of scaling data ingest and Machine learning at scale, and how Comcast has tackled these challenges to build a unified analytics and AI platform.

Why Voice is Important

Voice is a game-changer in the world of customer experience. It allows us to invert the interface, so instead of an interface that dictates to You what you can do, you can express what you want to do. This is a significant shift in the way we Interact with technology, and it has the potential to revolutionize the way we live our lives. Voice allows us to change the way our customers communicate with our devices and their homes, making it easier and more intuitive than ever before.

The Rise of Voice at Comcast

The rapid rise of voice at Comcast is best illustrated by the voice remote. Since its launch in 2015, the voice remote has been one of the most popular features, with more than 20 million voice remotes in the field today. Customers used it more than 9 billion times last year alone, and its usage is rapidly rising. The voice remote allows customers to describe what they want to do using their voice, and it's up to Comcast to figure out how to deal with that.

How the Voice Remote Works

The voice remote works by taking a query from the customer and sending it to the cloud to do automatic speech recognition. It then converts the query into a textual query and applies a number of machine learning algorithms to understand the intent and figure out the best action to take. For example, if a customer wants to know the latest news about artificial intelligence, they can simply speak into the voice remote and say, "Show me the news about artificial intelligence." The voice remote will then display a ranked list of the latest news about artificial intelligence.

Challenges in Scaling Data Ingest

One of the biggest challenges in building a platform for data analytics and ML is scaling data ingest. Comcast had to deal with ingesting billions of voice Sessions, petabytes of content, and telemetry data while handling millions of transactions per Second. This required a multi-hop enrichment process that combined hundreds of different event types into a flexible schema.

Tackling the Challenge of Scaling Data Ingest

To tackle the challenge of scaling data ingest, Comcast had to build a pipeline that could handle petabytes of data while being reliable, performant, and easy to use. They used Delta Lake to manage their data and built a pipeline that could handle streaming data, optimize small files, and scale to 64 machines. This allowed them to build a rich data set that could be used to make more informed decisions.

Machine Learning at Scale

Machine learning at scale is another challenge that Comcast had to tackle. They had to deal with hundreds of machine learning problems and manage hundreds of models that needed to be deployed into a disjointed set of environments in production. They also had to deal with the fact that machine learning models behave differently in production than they do during training.

Pain Points in the Machine Learning Lifecycle

The machine learning lifecycle is difficult to manage because data scientists and researchers want to use many different kinds of tools, generate hundreds of models, and collaborate with teams all over the world. They also need to deploy these models into a disjointed set of environments in production and monitor them to ensure they are performing as expected.

Developing and Managing Machine Learning Models

To develop and manage machine learning models, Comcast used MLflow to manage their models and Kubernetes to standardize their deployment environment. This allowed them to reduce their deployment time from weeks to under five minutes and gave them the ability to quickly iterate on their models.

The Big Void Between Dev and Production

The big void between dev and production is a significant challenge in the world of machine learning. Researchers want to use the latest and greatest tools, while production teams need infrastructure that is proven and highly performant. Comcast used a model abstraction to act as an interface between the developers and the production systems, allowing them to deploy models quickly and efficiently.

A Unified Analytics and AI Platform

Comcast's unified analytics and AI platform gives them the ability to improve reliability, be more productive, iterate faster, and make their customers happy. It allows them to build a rich data set, manage their machine learning models, and deploy them quickly and efficiently.

Conclusion

In conclusion, Comcast has built a platform for data analytics and ML that supports its mission of reinventing the customer experience with voice and AI. They have tackled the challenges of scaling data ingest and machine learning at scale and have built a unified analytics and AI platform that allows them to be more productive, iterate faster, and make their customers happy.