CIQ adds Service Endpoints to Fuzzball for AI control

CIQ has added Service Endpoints to its Fuzzball platform, extending the product from a workflow tool into a combined environment for AI training and model serving that runs in on-premises, cloud and hybrid set-ups.

The company said Service Endpoints allows organisations to run persistent services alongside batch jobs within the same workflow definition. CIQ framed the change as a way to keep model development and deployment within one pipeline, rather than treating deployment as a separate system.

Single workflow

Fuzzball already focused on orchestrating work across compute resources. CIQ said the new capability adds an option for teams to define service-style endpoints that stay running while other pipeline steps execute. The same workflow can include training, fine-tuning, validation and inference stages, according to the company.

CIQ positioned the release around "sovereign AI", a term used by vendors and governments to describe systems that keep data and model operations under the control of a specific organisation or jurisdiction. CIQ said customers can develop and serve AI models from on-premises or hybrid environments. The company said this reduces reliance on external platforms and keeps proprietary data within controlled environments.

"Organisations want to use AI with their proprietary data in environments that they control, without sending their data to external platforms," said Jonathon Anderson, Fuzzball Product Lead, CIQ.

Anderson described how CIQ expects users to structure work in the platform.

"Fuzzball Service Endpoints treats the entire AI stack as a composable unified workflow that can be iterated on and refactored to meet your specific requirements. This gives you complete control over your AI experience, whether you're fine-tuning a model or orchestrating a suite of coordinating agents," said Anderson.

Training and inference

CIQ said Service Endpoints combines batch computing for training and fine-tuning with persistent services for inference and interactive development. The company said teams can run endpoints alongside other workflow steps. It also said the approach reduces manual handoffs between development and deployment stages.

The company argued that organisations often face a choice between using cloud services for convenience or building separate systems to keep data and operations under tighter control. CIQ said Fuzzball workflows run in the same way across on-premises and cloud environments, including hybrid configurations.

CIQ also said the system defines and manages batch jobs, internal services and interactive endpoints inside a single workflow. The company said this keeps workflows portable across Fuzzball environments and reduces the need to coordinate multiple platforms.

HPC use cases

CIQ linked the Service Endpoints feature set to wider high-performance computing workflows, beyond AI development. The company said the platform includes native support for Jupyter, VDI and visualisation services. It also said users can inspect running workflows in real time and make adjustments while computations run.

CIQ said this model suits researchers who want to observe simulations in progress and validate results before jobs complete. The company said the same approach applies to interactive access patterns for data scientists and engineers who prefer notebook or desktop-style interfaces on shared compute resources.

Alongside sovereign AI, CIQ outlined two other categories it expects users to pursue: inference processes that operate like internal services, and interactive use of cluster resources. CIQ said traditional microservice platforms prioritise web functionality and can impose trade-offs for performance-oriented workloads. It said Fuzzball Service Endpoints supports coordination within and between workflows, including communications between batch jobs and persistent services such as databases and APIs.