PortOS
When fog, late ships, or crane breakdowns hit a port, cascading delays across berths, cranes, and union shifts cost millions.
Core trick
Argon proposes berth plans and playbook changes, and a CP-SAT solver's exact optimum verifies each change before it is kept.
Description
When fog, late ships, or crane breakdowns hit a port, cascading delays across berths, cranes, and union shifts cost millions. We built a self-optimizing 3D digital twin of the Port of Oakland's 4 terminals where a Gemini 4 Argon Player and Coach loop recovers disrupted schedules and runs live "what-if" scenarios, cutting unnecessary ship relocations by 67%.
README excerpt
# Port of Oakland digital twin, operated by a self-improving Gemini Argon agent Running a port is constant optimization. Ships arrive late, fog closes the channel, a crane breaks, and the berth plan has to be redone. A digital twin can run the port in simulation and help the people who operate the real one. This repository is a digital twin of the Port of Oakland, the container port across the bay from San Francisco, with its real terminals, cranes and shifts and an AIS-shaped ship feed. Gemini 4 Argon operates it: it makes the berth plans, answers questions about the port, makes schedules when asked, and improves its own playbook. It does not get to grade itself. Every change to the playbook is verified by code against the exact optimum from a CP-SAT solver. ## What you see The main screen is the 3D port: ships come in from the Golden Gate, wait at anchor, berth on their quay sections and get worked by quay cranes hour by hour, with late ships, closures and fog windows marked. The side panel has two tabs. Operator shows the port at the current hour (the quay to scale with each ship's sections and cranes, ships alongside, ships waiting at anchor, the next arrival, cranes working of the pool, late ship-hours, and one sentence naming the exception in force, such as a storm closure or a late ship) and a chat with the agent, which can explain a late ship, regrade a what-if with the real checker, or propose a plan and show it in 3D. Results has the "Optimize my playbook" button