System design for physical AI: a port digital twin that buys no hardware and owns every record

This engineering summary details a novel reference architecture for a digital twin system in the port industry. The system is designed to be completely software-based, requiring no additional hardware purchases. It operates entirely within the port's existing infrastructure.

The system comprises a class-based operator, not a named individual, and governs a comprehensive data model of the port's various assets and activities. This includes information about piers, berths, channels, parcels, surveys, inspections, and financial records. All of this data is collected and integrated through event-driven adapters, which feed read-only data into the twin.

The twin itself is based on a single, open embedding model for document retrieval, which processes a variety of documents such as drawings, inspections, and messages. This model is lightweight, weighing only a few hundred megabytes, and runs on existing enterprise virtual machines, requiring no additional GPU clusters.

The key aspect of this system is that it does not interfere with the port's existing operations. It reads data from the port community system without collapsing the line between the landlord and the operator. The twin provides insights and generates worklists for utility gaps and inspections, but pilots, berth planners, engineers, and financial staff continue to use their own systems of record.

From a cost perspective, the system is virtually free beyond the integration and software costs, which are already accounted for in the port's existing expenditures. This makes the system an economically viable solution for modernizing port operations without significant capital investment. The design and object model are available for download, allowing other ports to implement similar systems.