by Massimiliano Pardini, CTO of Net7
The call comes in August
There is a scene anyone who works in yachting knows well. A boat at anchor in the height of the season, unable to carry on with the holiday. An onboard system that has suddenly stopped. And a technician, hundreds of kilometres away, trying to work out over the phone what is happening from a rough description.
There is a great deal of talk about AI in shipyards and very little to show for it, and the most useful contribution I can make is to describe a case that works: what was there before, what is there now, and what difference it makes to the people doing the work.
What happened before
The case I follow most closely concerns some of the onboard systems of superyachts, which we are working on together with a company specialising in the sector. Systems nobody notices as long as they work, and which, when they stop, bring life on board to a halt, almost always far from the shipyard that built the vessel.
Until recently, a fault was discovered once it had already happened. Diagnosis depended on the technician who came on board, and that kind of experience is concentrated in a handful of people the industry knows by name. Manuals, diagrams and service history were scattered across folders, emails and the memory of whoever was there that time. The paradox is that the data was already there: sensors and control units have been recording for years how those systems work, minute by minute. It stayed inside the machine, and nobody looked at it until something broke.
What changes today
In the project we are carrying forward, those signals are read continuously to spot when a component drifts away from its normal behaviour: a value that slips slightly with each cycle, a machine that takes a few seconds longer to do the same job, things invisible to the eye and obvious over weeks of data. Alongside the signals is the second half of the project: a knowledge base that brings together manuals, diagrams and past interventions, and that AI makes searchable with a question, so that the alert arrives together with what to do.
The real change is in the calendar more than in the technology. A planned intervention during the refit window costs a fraction of an unexpected breakdown at the quay in August. And the experience of the technician who knew where to put his hands no longer retires with him: it stays in the system and grows with every intervention.
I find the same pattern, in a far less spectacular form and already in production, in shipyards.
With NaviSafe, the document compliance platform for shipyards, the problem lies in hundreds of PDFs: DURC certificates of social security compliance, fitness-for-work assessments and certifications for every supplier on site, managed by hand across emails and shared folders, with the risk of a worker entering the yard with an expired record.
Here artificial intelligence does two precise things. It recognises whether the uploaded document really is the one required, by comparing it with the template, and extracts the dates and data needed. These are the points where traditional OCR and hand-written rules give up: every authority has its own form, every scan its own quality, and no definition holds for long. From there on, verification is automatic, cross-checked against the turnstiles, with an alert before expiry.
In both cases AI does a job that previously nobody had the time to do: read everything, all the time, and flag in good time what doesn’t add up.
Ideas remain our job
A few days ago I wrote that we are delegating a growing share of our work to agents, and that the ideas and improvements to bring to projects remain our job. It is the second question of the interview, and here I am talking about the trade I know best: ours, the trade of people who make software.
Until a year ago, people said coding was largely solved. The phrase is already dated: today coding is solved entirely. Giving a model a well-defined problem and getting working code back has become the easy part. A developer’s value, measured for thirty years by the ability to write that code, has moved elsewhere.
To two places, to be precise. The first is the architectural phase: deciding how the system is built, which pieces talk to which, where the boundaries lie, what must stay simple. These are decisions that generated code does not make on its own and which, if wrong, no speed of generation can make up for. At Net7 we work this way, with OpenSpec, an open-source spec-driven development framework: first we write what the system must do and how it is structured, then the AI works against that specification, which remains readable and verifiable by a person.
The second place is context. A model works as well as the project allows it to: if it knows the conventions, the folder structure, the patterns already in use and why they were chosen, the result reads like the rest. Otherwise it works, and nobody wants to touch it. Building the right context has become a real craft, and it counts twice over: what comes out must remain readable by those who come after us and by the next agent that gets its hands on it.
Readable, though, is a word that says little on its own. Danilo, who wrote about the copywriter’s craft in these pages, explained it better than I could: for a text, the right question is whether the reader, afterwards, knows what to do. For what we produce with AI the criterion is the same: will whoever opens it six months from now, person or model, understand what it does and why? If the answer is yes, the job has been done well. If not, it is a disaster waiting to happen.
Those who know the foundations
The foundations, in yachting, are the people who recognise the wrong noise from a machine before any sensor does. In software, they are those who learnt by making mistakes, back when nobody could yet ask a model.
It is these people who today carry an extra responsibility: putting their experience into the specifications, and passing it on to those who come after. That is why I still make a point of saying that I started out as a developer, before becoming a PM and CTO. Those who have travelled that road can tell, in a control unit as in a repository, the signal that matters from the noise. That is something AI still cannot do on its own. It reads the data that was already there, on board and in the code, and does so better and faster than we do. Deciding what to do with it, and teaching that to those who come after, remains our job.
Massimiliano Pardini is CTO of Net7. For the maritime and nautical sector, Net7 develops digital platforms for shipyards, marinas and port operators. These include refit order and naval project management with In.Refit and DR&AM, shipyard safety and document compliance with Navisafe, and access to port services with Ormeggiando.
