In this article, I’ll shed a light on what are the ready-made configurable components in the area of Cloud AI & MLOps that we have built to speed up AI development with our customers. Configurable components tackle the main challenges for large enterprises: 1) sustainability of AI R&D, 2) scalability of AI processes, and 3) growth of AI R&D.
In addition to a team of more than 180 leading AI experts, Silo AI provides tooling and modular AI solutions and components with Silo Operating Software (Silo OS). Silo OS is designed to significantly reduce the risk, cost, and time-to-production in AI product and solution development. Cloud AI & MLOps is one of three focus areas of Silo OS – others being Embedded AI and the Internet of Things.
Our offering regarding cloud and MLOps tackles the three biggest challenges for enterprises:
- Sustainability of AI research and development: Keep track of and govern work products (data, code, models) to make sure that the quality standards are met. Build and manage the AI platform and pipelines and monitor systems. Deploy AI/ML models to optimize costs and improve results.
- Scalability of AI processes and products: Recycle work products and apply automation and deploy AI functionality in standardized ways to increase results and to support multiple cases.
- Refactoring and growth of AI research and development: Build AI on cloud while balancing lock-in and investments and build self-service for internal developers and external partners to grow AI R&D.
Accelerating the building and adoption of MLOps with configurable components
At Silo AI, we have worked with hundreds of real-life AI projects to bring machine learning solutions from the R&D initiatives to production. We have witnessed firsthand how companies struggle with organizing themselves around AI projects, experienced the pain of using nonoptimal tool stacks (f.ex. tools meant for DevOps bent for MLOps purposes), and trying to build quality models without transparency or ability to reproduce the previous training runs easily and reliably.
There isn’t such a thing as off-the-shelf MLOps tooling that companies can just take, set up, and start using. Even the machine learning platforms as a service (ML PaaS) offered by cloud vendors – like AWS SageMaker, Azure ML, Google’s Vertex.ai, and Databricks – need configuring and adapting to the company-specific needs and not to mention cloud-agnostic machine learning setups – Kubernetes + open-source MLOps technologies – that require even more care. Some cases require combining cloud-specific PaaS with cloud-agnostic components because the PaaS may cover a lot of the requirements but may not resolve all the needs.
After mapping the organization-specific processes for the AI development needs, the toolstack needs to be selected and configured to suit those needs. Silo AI is committed to offering our clients assets for configuration that takes companies a step closer to the production-ready infrastructure right off the bat. This is why we decided to build our own configurable components that can be plugged in according to the AI R&D needs. Instead of aiming for a one-size-fits-all solution, we want to embrace the full complexity of production-grade machine learning projects that need to take into consideration the DataOps, MLOps, and DevOps.
Building technology-agnostic Cloud AI & MLOps components is one of the core values driving our development work. As a consulting company, we have the privilege to work with top-notch companies that all have differentiating needs and tool stacks. From this experience, we have gained knowledge from open source and cloud machine learning offerings and an understanding of the special needs for each.
Cloud AI and MLOps components
Below I’ll describe our MLOps related component offering in brief and if you are interested in discussing any of these further, you can contact me directly.
Annotator for DataOps needs
Managing data annotation and labeling operations at scale is one of the critical activities of a successful machine learning project. Our Annotator tooling is battle-tested in several projects; it’s flexible and customizable for your data formats, system integrations, domain expert workflows, and use cases (computer vision, NLP, signals, etc.). With the Python SDK or the REST API, our Annotator can be fully integrated into your system. We provide build-in UIs for image, text, and time series annotation but the UI can also be customized for your use case.
Production-grade pipelines for MLOps core and DevOps needs
As previously mentioned, whatever MLOps tool stack you choose, it always requires configuration. We have built production-grade pipelines that provide critical ML development and AI deployment functionality in all popular cloud environments. It’s built for job orchestration to bring continuous integration and development practices and automation to machine learning development. The production-grade pipeline templates enable you to stitch data preprocessing, model training, and validation steps together with version control and transparency for the model training. Also, you can create your own custom way for setting up and running pipelines in standardized ways across teams and organization and from research experiments to production-stage runs.
Production-grade MLOps platform for platform needs
Our production-grade platform provides microservice setup software for running data, machine learning and DevOps pipelines in all popular cloud environments.
Depending on your technology and cloud choices, there can be restrictions and other considerations regarding specific AI R&D requirements of a large enterprise. ML PaaS in AWS, Azure and GCP may not cover all the functionalities required of the MLOps platform; They could be too costly for certain use cases, or in general, there could be a need to move between clouds for the sake of agility. Our platform offering and its microservices – for e.g. versioning, pipeline orchestration, and monitoring – can be used in all cloud environments (e.g. AWS, Azure, GCP) that provide managed Kubernetes environments as infrastructure as a service (IaaS) and can also be moved between the clouds thus providing cloud-agnostic MLOps stack which is something that we have already done with our clients.
The public cloud might be out of the question for certain contexts like sensitive or big data in healthcare or a need for air-gapped systems for security purposes. With our platform – building on top of Kubernetes and popular, well-supported open-source MLOps, DevOps, and DataOps technologies – we also meet these kinds of extra requirements in the MLOps space. This way we can bring our MLOps services also to the most sensitive business cases and enable standardization of MLOps workflows, automation, and the AI R&D scaling benefits for those.
MLOps needs addressing
Getting back to the three most pressing challenges that we have seen our customers face in building their MLOps capabilities:
- Sustainability of AI R&D
- Scalability of AI processes and products
- Refactoring and growth of AI R&D
Our product offering standardizes the basis for AI development ensuring production-level quality and effective operations, which in turn, makes AI development sustainable. We offer and enable you to store reusable assets, like quality-controlled data in feature stores and governed model versions to model stores, to scale AI development to speed up time-to-market and increase return-of-investment. The wide cloud platform coverage of our Cloud AI & MLOps product offering brings more flexibility and room for growth.
To learn more about Silo Operation Software and our configurable components in the area of Cloud AI & MLOps to speed up your AI adoption, contact me.
Jukka Remes, PhD, Lead AI Solution Architect, Tech Lead / Cloud AI & MLOps
