Jukka Remes describes an open-source Kubernetes platform that packages Kubeflow, MLflow, KServe, Prometheus, and Grafana for deployment across laptops, virtual machines, and supercomputers.
2
The platform gives researchers reusable ML pipelines, Git-based project repositories, and CI/CD automation so they can focus on model work instead of manually operating infrastructure.
3
Jukka's 8wave AI product connects business requirements and AI regulations with ML operations, recording requirements, lineage, performance, and evidence for production and audit decisions.
Summary
Jukka Remes explains how a platform developed at Silo AI became an open-source setup for researchers, engineers, and organizations that need MLOps without relying on a cloud provider. It packages several open-source components into an installable Kubernetes environment and supports workflows from training through deployment and monitoring. Researchers can use shared configuration repositories, project repositories, pipeline templates, and CI/CD automation while keeping their work in Git. Jukka connects this approach to reproducible science, since reviewers and collaborators can inspect experiment code, parameters, and pipelines. He also discusses the gap between research and product development, especially as AI regulation creates a need to trace how models were built and operated. Through 8wave AI, he is building a management layer that connects business requirements and regulatory expectations to ML operations, so teams can check whether a model meets the purpose and conditions agreed by its stakeholders.
The platform grew from practical client work at Silo AI
Jukka Remes joined Silo AI as a solutions architect in 2020 and worked with clients that needed MLOps practices and platform setups. His team combined open-source components into a loosely integrated platform for cases where Silo AI managed model development and deployment. More mature clients could also reuse selected parts of the setup in their own platforms. The initial scope covered pipeline orchestration, model and metadata management, experiment tracking, deployment, and some monitoring. Data management was considered, but the team did not ultimately extend the platform in that direction.
Open sourcing made the platform useful beyond its original consultancy setting
Changes in Silo AI's research direction created an opportunity to release the platform as open source. The University of Helsinki and other consortium participants were interested in continuing the work. The university extended it for high-performance computing and Finnish supercomputing centres, and Jukka continued developing it with students at Haaga-Helia. The platform is intended to cover many kinds of ML work because its Kubeflow and MLflow foundation is generic. Jukka is less certain about how efficiently the current components handle large-model training and says that this is an area he is exploring with the University of Helsinki.
Reproducible pipelines can reduce the administrative burden of research
Jukka draws on his earlier work with brain-imaging analytics, where researchers had to run many combinations of parameters, filters, and data preparation steps. Managing those experiments manually was tedious. A platform can bring consistency to such complicated processes and make the results easier to trust and review. Researchers can share experiment setups, code, and parameters, although data-sharing restrictions may still apply. Sharing a complete pipeline gives reviewers or other researchers a way to reproduce the work more directly.
A shared platform can narrow the gap between research and product development
Jukka says research results are often separated from product development, including inside large companies with dedicated research centres. Moving a result into a product can involve a substantial handoff, even when the research is promising. He expects AI regulation to increase the need for records that begin during research rather than being reconstructed later in product development. If teams must explain how a model was developed and operated, lineage and experiment history need to exist across the full process.
The platform packages existing tools instead of hiding them behind a large custom system
The setup combines Kubeflow, MLflow, KServe, Prometheus, and Grafana as an integrated package. Jukka describes the services as still fairly loosely coupled. The value is that users can install the package without following a long set of manual or custom infrastructure instructions. The main open-source installer runs Kubernetes in Docker through kind, which lets users provision the platform on their own computer or on a machine in a supercomputing centre. A separate managed Kubernetes deployment path exists for Google Cloud.
Git repositories and CI/CD move platform operations away from the ML developer
Students working with Jukka built tooling around Git repositories and CI/CD pipelines. A configuration repository can hold shared platform details for a team, while individual ML project repositories link back to it. The tooling can create the repository structure and CI/CD workflows. When a developer commits to the relevant Git branch, the pipeline is compiled inside CI/CD and deployed to the connected platform instance. This lets researchers contribute Python code to pipeline steps without manually handling platform API calls or deployment procedures.
The open-source setup supports research and production workflows, with examples still limited
Jukka says the project currently has example pipelines for training and deployment. Users can also run those stages separately. The platform can be installed on a local computer, a virtual machine, or another available environment, so it does not require access to a cloud service. He describes the software development and operations layer as more broadly reusable than the specific Kubeflow and MLflow installation. The current examples are a starting point rather than a complete library of ready-made workflows.
8wave AI adds business and compliance requirements to ML operations
At 8wave AI, Jukka is building a digital product that connects AI operations with the business purpose of an AI project and with regulatory expectations. Users can record requirements on the platform and follow how those expectations are met during ML operations. The product tracks what happened so teams can understand lineage, performance, and whether targets were met. Jukka says projects often fail because teams work in separate organisational corners and build toward the wrong business need. His product aims to give stakeholders a shared view without requiring them to create and maintain a custom system.
"You can put the expectations for things on our platform, and then you can take it to the ML operations side and follow up on how those expectations are met."Jukka Remes43:43
Who should watch
You are building ML systems with researchers who need a simpler path from experiments to deployment, especially when a managed cloud platform is unavailable.
Your team has repeated experiments manually and needs shared pipelines, Git repositories, or CI/CD without assembling every infrastructure component itself.
You need to connect business requirements and AI regulation with model lineage and operational evidence instead of collecting that information after development.
