Private AI aboard a small vessel
A project plan, not a case study: what it would genuinely take to run useful local computing on a boat, given the power budget, the heat, and a connection that disappears.
Read this first
Nothing here has been built yet. This page exists because the plan is worth publishing and because a prospective client deserves to see how I scope something before there is a result to show. When there is a prototype and measured numbers, this page changes and the date on it changes with it.
The constraint set
A boat is a genuinely hard site and an excellent honesty test. The house bank is finite and shared with things that matter more than compute. Heat has nowhere convenient to go. Connectivity is intermittent even with a good satellite terminal. Vibration and salt air are real. Nobody qualified is coming out to the mooring to reseat a card.
Where local computing might actually win
Not chat for its own sake. The candidates are the boring ones: searching manuals, schematics, and maintenance history without a connection; document and inventory work; local services that stay available when the link drops. These are non-safety-critical by design.
Where it will not go
No navigation advice. No integration into propulsion, steering, or safety systems. No autonomous operation. Those boundaries are not caution for its own sake; they are what separates a useful tool aboard from a liability, and they will stay in place regardless of what the hardware turns out to be capable of.
What would have to be measured
Draw at idle and under load against the actual house bank. Thermal behaviour in an enclosed space in summer. Time to useful answer on a defined corpus with a stated model and settings. Behaviour when the link drops mid-task. None of these numbers exist yet, and I will not publish estimates dressed as results.