Towards a Hybrid AI reference architecture
It’s a whole new world, setting new challenges of definition, integration, and coordination of technologies and activities, and that’s why at Capgemini Engineering we believe a reference architecture is needed. Such an architecture should describe a standard approach for developing AI solutions, capturing a common understanding of key AI, Gen AI, and Hybrid AI concepts in an engineering context.
AI will shape and integrate with all aspects of engineering – the products that we build and the processes and facilities that we use to create them. To ensure that we develop consistent and efficient solutions, in our products, our design offices and our factories, we need consistent language, consistent structure and clear vision of the interfaces and roles of components – provided by a reference architecture.
Engineered products and engineering applications face constraints and expectations which make the AI landscape of engineering complex and constrained. A reference architecture must address functionality, connectivity, and data distributed across a range of physical and virtual environments – from hyperscaler data centers through network operator and enterprise facilities, to individual sites, equipment and portable devices. These will have different computational and physical constraints: of computation, storage, connectivity (bandwidth and latency), size, weight and electrical power.
The Hybrid AI realizations will be similarly varied: from interfaces and lightweight inference in embedded devices, through federated learning and private knowledge bases to foundation models and their associated training infrastructure.
Architectural decisions will need to consider data sharing and distribution, privacy and security, the compute and storage facilities required, communication and connectivity, and load distribution.
The architecture also needs to address regulatory, legal, and safety frameworks.
Hybrid AI provides new tools to navigate this landscape and new options for creating engineered products and processes. The combination of generative AI and structured knowledge means that we can distribute information and activity across complex environments in ways that are unachievable by legacy approaches or monolithic generative solutions.
A set of principles like these will bring consistency, not only to the terminology used by engineering organizations, but to the approach they take. It will mean that, while Augmented Engineering solutions can be thought of custom 'vehicles', bespoke to the needs of the organization that deploys them, the wheels on which these 'vehicles' run won’t need constantly to be reinvented.
