This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/port-digital-twin-us/. The operator is described by class, never by name.
A landlord port authority in the United States owns the piers, berths, channels and parcels but not the terminal operating systems that run on them. Its critical information sits in disconnected systems: the port community system, the GIS estate, AIS vessel positions, truck gate transactions, bathymetric surveys, subsurface utility maps, inspections, engineering drawings and the finance backbone. The ask is one governed twin of all of it, inside the continental United States, on infrastructure the port already runs.
Here is how that turns into a system with nothing to buy.
1. Join first, model second
Every source enters through a read-only adapter onto an event backbone, and the adapters write thirteen typed objects: berth, navigation channel, subsurface utility layer, environmental sensor feed, parcel or facility, bathymetric survey surface, utility gap record, capital project, inspection record, vessel movement, truck movement, PCS message and engineering document.
The berth is the focal object. From one berth a single traversal reaches the vessel movements that touch it, the channel it adjoins, the current bathymetric surface on that channel, and so the under-keel condition a pilot needs. From a parcel it reaches the lease, the revenue and expense lines, the capital projects whose footprint overlaps it and the inspections against its facilities. A GIS layer shows where things are; the ontology says how they relate.
The object model ships as JSON (ontology/objects.json, format hyper-ontology/1) so you can load it instead of redrawing it.
2. One model, sized honestly
The only model in the register is BGE-M3, an open embedding model for document retrieval across drawings, inspections and messages.
weights ~2.27 GB at float32, ~1.1 GB at fp16, ~0.6 GB at 8 bit
KV cache none: an embedding model generates no token stream
hardware none bought: existing enterprise virtual machines
There is no GPU cluster to size because the requirement asks for none. A hosted document AI API was weighed and set aside: it would move port documents outside the boundary the port's own security requirements set.
3. Keep the boundary and the authority-operator line
The twin runs on port-managed servers under the port's own change management, inside the continental United States. It reads terminal operators' data through the port community system and never collapses the landlord-operator boundary: the port sees, it does not run the terminals.
4. Nothing writes back
The twin shows. Pilots, berth planners, engineers and finance staff act in their own systems of record. The only records the twin writes are its own: utility gap records and inspection worklists, each anchored to the asset it concerns.
5. What it costs
There is no hardware line. The weights fit a standard virtual machine, and the requirement buys no cameras, sensors or servers. The cost is integration and software on infrastructure the port already pays for, which is why the paper carries no cloud comparison.
Full design, figures and the object model: the paper.
Designed on CodeNinja Praxis, the platform for designing physical AI systems. The object model imports into Hyper Ontology, which turns it into a living system. Load it yourself with the open hyper-ontology loader.
Top comments (0)