I spent enough watches on a tanker bridge to know what fails. The decision at 0300 should not depend on whether the chief got sleep. The open source frontier models are good enough now. We take them and we run them on the ship itself. No cloud round trip. No waiting on a sat link. The model watches the things a tired master would miss, round the clock.
We are master mariners, naval architects, and engineers. We have run tanker fleets, built crewing systems, and shipped open source for voyage planning, bunker optimization, and inspection compliance.
A tanker at sea is a long way from a data center. Bandwidth is thin and it drops out. So we run the model locally on board. It reads the noon report, the bunker prices, the maintenance log, and the inspection history. It works when the link is down. It works in the middle of the Indian Ocean.
We are not building an autonomous ship. The master runs the ship. The co-pilot reads the same signals the master reads and points at the two or three things worth a second look this watch. It shows the reasoning. It does not sleep. It does not forget the handover.
Every fleet is different. A 15 year old Aframax on a clean trade is not a new VLCC on the Gulf to China run. The system adapts to your vessels, your crews, and your trades. It learns what your best fleet manager would have remembered.
The kit on a tanker has to work for years in salt air and vibration. We build it the same way. No shortcuts. The crew and the shore office have to trust it at 0300. If they cannot trust it they will stop using it. So reliability comes first.
We pull together the noon report, the sensor feed, the inspection photo, and the crew report. One picture of the ship, not five spreadsheets. That is how you catch a bearing before it fails or a voyage plan that burns a tonne too much.
We run small pilots on real ships. Each one has a clear goal. A maintenance flag that catches a failure before it grounds the ship. A voyage suggestion that saves fuel. A crew shortlist that holds the line on vetting. The ones that prove out become services. The rest get written up and shelved.
Safety is not a slide. We keep a high bar on what the released system can do and we give the operator the freedom to use it. We share what we learn about building safe operational AI with the rest of the industry. We share code, datasets, and model specs so others can test it. Red teaming and post deployment monitoring on today's ships teach us how to build tomorrow's.
We measure what matters. Fuel saved. Failures caught early. Crew retained. Not a dashboard metric. Real value on a real voyage.
If you operate tankers and have a problem worth solving, we will run a pilot with you. We bring the models and the engineers. You bring the ship and the problem.