Somewhat ironically, productionising Machine Learning solutions continues to be quite a manual discipline. Data scientists and machine learning engineers take care of tasks that they perhaps shouldn’t have to and spend less time doing what they are meant to, leading to inefficiencies.
But it doesn’t have to be this way – and if we at Vodafone have our way, it won’t be. As the solution architect for our AI Booster programme, I’d like to tell you all about how our innovative Unified ML Platform could be a complete game-changer.
Vodafone’s AI Booster Project is a unified machine learning platform we’re building on top of Google Cloud Platform (GCP). The AI Booster programme captures all aspects of the machine learning journey: data validation, data preparation, feature generation, training, prediction and finally, operations and monitoring.
There are a wealth of different roles across a machine learning programme – Data Scientists, ML Engineers, DevOps professionals, Data Engineers and so on – and each tends to stay within their silo. However, there can be a lot of technical debt as a result. As such, our AI Booster attempts to better harmonise each discipline via standard templates while also automating much of the integration and deployment. People can focus properly on their actual work without wasting time on mundane, repetitive tasks like how to configure bastian hosts, proxies and connectivity between different services.
Additionally, the AI Booster platform is fully cloud native, utilising GCP’s serverless components – hence, it’s extremely cost-effective. We pay for compute resources as we go and don’t have to worry about server management. This makes it super scalable too – we can experiment with small use cases, and then industrialise them in no time without any capacity planning constraints.
Overall, AI Booster is a template solution that we are offering to our Data Science community as a fully packaged set of products and services deployed on GCP. Requesters need only fill out a form to request the service, and we provide them with a Platform as a Service with Vertex-AI at its core.
Ultimately, what we’re trying to do at Vodafone is unify our machine learning efforts using a standard template. We’re in an enviable position in that we have multiple machine learning platforms, programmes and processes running in parallel. But while exciting, this can lead to inefficiencies including duplications of activities and artefacts.
To solve this problem, we’ve introduced machine learning operations, or MLOps. Already across Vodafone we have slowly replaced monolith applications with microservices and now we hope to do the same thing in machine learning.
MLOps unifies our ML initiatives to avoid duplication and enable us to work faster and more cost-effectively. In short, our aim now is to build reusable components in smaller chunks, which can be spread across different use cases. Reusability doesn’t just stop at MLOps in AI Booster -our Platform components and shared services also follow this principle. We tend to build solutions that are generic, scalable and can accommodate a wide number of use cases.
It can almost be thought of as an assembly line on a factory floor – everyone works within their specialty at different points on the conveyor belt, and by the end it’s all been pieced together into a complete and fully functional solution. All the different ML stages are captured and it offers a solid foundation to work from and framework to work within.
Since the start of this programme, our costs and operational overheads have gone down drastically, in large part due to replacing our self-managed servers with the GCP-managed products. We’ve drastically reduced solution deployment times as well, by creating reusable and generic Vertex Components and CICD pipelines; previously we’d face endless manual steps, but we now simply push code to the GitHub repository and our automation takes care of the rest.
Additionally, we’ve centralised security scanning and governance for the entire set of AI booster use cases, which gives us more control and flexibility and greater security. We make sure that our developers are working in secure environments where our data is optimally protected.
But automation is perhaps the biggest benefit. It makes such a difference that we’ve actually made the AI Booster available as an API. Our backend automation has abstracted things so much that all you have to do is send an API request with your configurations and it will take care of all of the platform set up.
Under the hood our backend manages a series of tasks we previously had to do manually: projects creation, Identity and Access management, Security Perimeter set up, and our Networking configuration, the most complex of all. As all services must be private for security reasons, Networking layer communication across services can be very challenging, but within AI Booster we have managed to automate everything so our customers aren’t even aware of it.
In recent years Vodafone has started to pivot from being a Telco to a technology company. We’re hiring a lot of Software, ML, Data and Platform engineers. We want to utilise open source code and resources, and we want to give back to these communities. AI Booster aims to unlock this for our engineers and hopefully as we grow, the quality software and components they build can be shared with the open source community as well as across Vodafone.
This is a simple and accessible platform that can be used to build machine learning solutions using a variety of different frameworks, including TensorFlow, XGBoost and even Google’s no code AutoMl. It allows engineers to focus purely on their activities, not the complex and time-consuming processes that surround them.
All in all, it’s been a dramatic change but it’s also been very rewarding for all of us at Vodafone. I’m really looking forward to where this journey is heading!