AI Sailing Analysis on Typon, a 1976 Swan 47

Short version: we put a Raspberry Pi on a fifty-year-old racing boat, recorded every instrument reading for a week of racing on San Francisco Bay, and pointed AI at the result. The AI sailing analysis that came back spent most of its effort on something we had not asked for. Before it could say one useful thing about how we sailed, it had to work out how far each instrument was off, and correct it.

Nothing was broken. The compass, the speed sensor and the wind vane were all doing their jobs. They were simply installed on a real boat, where a compass sits near metal, a paddle wheel picks up weed, and a wind vane at the top of a mast is never quite aligned with the hull below it. That is configuration, and configuration drifts. Anyone who has run a monitoring system knows the feeling.

What follows is how we got the data off the boat, what the analysis found, the one thing it honestly could not settle, and the part that turned out to matter more than any of it.

A Raspberry Pi on a Fifty-Year-Old Boat

Typon is a Swan 47, built in 1976 and raced by a crew of fourteen under her captain, Peter Rostas. In September she raced the Rolex Big Boat Series on San Francisco Bay, six races over five days out of the St. Francis Yacht Club. The regatta is in its sixty-second year, ran 81 boats across eight classes, and doubled as the ORC North American Championship. Typon raced in ORC Class D.

The hull is from 1976. The instruments are not: they have been refitted over the years, which is the only reason any of this was possible. Modern marine instruments talk to each other on a shared bus, and anything else plugged into that bus can listen to the whole conversation. So we plugged in a Raspberry Pi.

Typon at the 2026 Rolex Big Boat Series. Photo: Rolex / Daniel Forster.

Recording a Week of Racing

The Pi did one job and did it without being asked twice. When the navigation instruments came on, it started recording, which matters more than it sounds. Nobody on a racing boat is going to remember to start a logger in the five minutes before a start, and a logger that misses the start is a logger that misses the interesting part.

It wrote down everything on the bus: position, speed over the ground and course from GPS once a second, compass heading and how far the boat was leaning ten to twenty times a second, speed through the water, the wind angle and speed at the masthead, rudder angle, depth. It also listened to the AIS transmissions of every other boat within range. Over the week that came to 84 log files.

The AIS part was the pleasant surprise. Boats broadcast their position and speed for collision avoidance, which means a boat racing next to you is continuously telling you where it is and how fast it is going. Not everyone in our class had it switched on, so the comparison is partial, but two boats we were racing against transmitted throughout and we have their tracks.

Three acronyms, briefly.

• NMEA is the language marine instruments speak to each other. One shared bus, every reading on it, and anything connected can listen.
• AIS is the transponder that makes a boat visible to others over VHF radio: transmitting position, speed, and course, every few seconds. Built for safety, useful here for something else entirely.
• An ORC certificate is a measured prediction of how fast a specific boat should sail at every wind speed and angle. It is what race handicaps are built from, and it doubles as a target. Score yourself against your own certificate, and the question becomes how well you sailed the boat you have.

First It Told Us the Instruments Needed Correcting

We wanted to know where we lost time. What we got first was a list of things to fix before that question could even be asked.

The compass. Typon’s hull is fibreglass, but a boat is still full of things that pull a compass around: the engine, the rigging, the electronics, the wiring behind the panel. Ours read about four degrees off on one of our usual headings and about a degree off on the other. Left uncorrected, that difference alone would have shown up as the boat slipping sideways two or three degrees on every tack, and we would have gone looking for a keel problem that does not exist.

The speed sensor. A small paddle wheel under the hull measures speed through the water. Ours read up to five percent low, and by how much drifts as weed and barnacles build up on the hull, so the correction is refitted for each day of racing. It also behaved differently depending on which way the boat was leaning, because a paddle wheel mounted slightly off the centreline sees disturbed water when that side of the boat lifts.

The wind vane. Two separate things, neither of them a fault. The vane is mounted a couple of degrees out of line with the boat, which is normal. And the sails bend the air before it reaches the top of the mast, by about five degrees in light wind and one and a half in strong. Every sailing boat has this. Almost nobody measures it.

The height of the mast. Our masthead is around twenty metres up. The ORC targets are quoted for wind measured at ten metres, where it blows slightly less hard. Nothing is wrong with either number; they are just not the same number. Compare them directly and every target you measure yourself against is the target for more wind than you actually had.

None of these are corrections anyone typed in. They are refitted from the race data every time the analysis runs, which means they stay honest as the boat changes.

Why Any of This Can Be Worked Out at All

The idea underneath the whole thing is simple enough to explain without mathematics. It is not a clever piece of engineering. It is a fact about sailing.

To sail upwind you zigzag, turning the bow through the wind each time. That turn is called a tack. Here is the fact: the wind does not care that you turned. Its direction is the same one second after the tack as one second before.

So any number that jumps when you tack is telling you about your instrument, not about the weather. Fit the wind vane corrections so that the calculated wind direction stops jumping across seventy-two tacks, and you have calibrated a masthead sensor without a wind tunnel, using nothing but the racing you were doing anyway.

The same logic does the rest. Sideways slip flips to the other side when you tack, while the tidal current does not, so comparing both tacks separates them: that came from forty-six upwind stretches. The compass was checked in the moments the boat was sailing nearly upright, because with no lean there is no sideways slip to confuse the picture. Each correction uses a situation where one unknown cancels itself out.

The One It Could Not Settle

Not everything came out clean, and the analysis says so in its own methodology rather than quietly rounding it away.

The wind speed at the masthead reads about six percent higher going downwind than going upwind. One of those two is right and the logs cannot say which. Settling it needs a flat calm day and a motoring test, which we did not have during a regatta.

The tidal current is the bigger caveat. San Francisco Bay has a lot of it, and we do not measure it at all. It is inferred from the gap between where the boat actually went and where its own instruments thought it was going. That works, but it comes from the same sensors as everything else, so it inherits their errors. It is an estimate, and the analysis labels it as one everywhere it appears.

An early attempt to fit the speed sensor by comparing it against GPS ran into exactly this. The two tacks are sailed in different water, with different current, so the answer changed with every choice of time window. That approach was dropped and replaced with one that does not need to know the current: over a steady minute the current is effectively constant, so how much the GPS speed moves for each knot on the paddle wheel tells you the sensor’s response without ever measuring the water.

Then It Told Us About Our Sailing

With the instruments corrected, every second of racing could be scored against Typon’s own certificate. Not against the fleet, and not against a handicap: against what this boat, measured, should be capable of in the wind we actually had.

Four words for the numbers below.

• Upwind and downwind: sailing towards the wind, in a zigzag because you cannot sail straight at it, or away from it. Different targets apply to each.
• VMG (Velocity Made Good) is how fast you are getting closer to where you want to go, not how fast the boat is moving to your heading.
• Heel is how much the boat leans. Leeway is the sideways slip that comes with it: the bow points one way, and the boat travels a couple of degrees to the side of that due to tide and wind.
• Percent of target is our VMG against the best our certificate says this boat can do in that wind.

The pattern across the week was consistent and not flattering. In light wind we sail at around 97 percent of target upwind. As the wind builds that falls away: 92 percent, then 87, and about 85 percent once it is properly blowing. Downwind is better throughout.

Underneath that is a specific habit rather than a vague slowness. We sail a few degrees wider than the certificate wants, and slightly faster than its target speed. That sounds harmless, and it is the whole problem: speed feels good at the wheel, and the boat is going quickly in a direction that is not quite where the next mark is. The correction is to point higher once the speed is already there, which is easy to write and hard to do at two in the afternoon with waves coming over the bow.

We are not going to pretend this made us fast. It told us where the time goes, which is a different and more useful thing, and what it mostly says is that we have a lot to improve and we simply need to sail better.

From Numbers to a Debrief

A table of percentages does not change how anyone sails. What does is being able to point at the moment it went wrong, so the output is a replay you scrub through. Our track and the rivals’ tracks are drawn on the chart, coloured from red to green by how close to target we were at that instant, with the wind, the angles and the targets in a panel beside it. There is one for each race.

Those tracks turn out to be worth more than lines on a chart. AIS gives a position, a speed and a course roughly every thirty seconds, which is thin next to our own once-a-second logging, and it is still enough to work with. Every boat racing under this rule has its own certificate, so a rival can be scored exactly the way we score ourselves: percent of what that particular boat should be capable of in that wind, not percent of what we should be.

That is a more useful comparison than who was ahead. Being behind a faster boat is normal and says nothing. Sailing at 85 percent of our own potential on a leg where another boat sailed at 95 percent of hers is a finding, and so is the reverse, which happened on the reaches. It tells us which parts of the course to work on rather than leaving us with a general sense of being slow.

It comes with a health warning the analysis states plainly wherever the figures appear: a rival’s percentage assumes it had our wind and our current, measured where we were rather than where it was, so a couple of points either way is noise. And we read it to find our own lost time. How another crew sailed is their business, not ours to publish.

Each race also gets a written review, and the most useful thing about it is how it ends. Not with conclusions, but with questions for the crew. Why did we hold that tack for six and a half minutes after the start? What happened at the first mark? What set our angle on the second beat, waves or trim or steering for speed? The logs can show that something was slow and usually where. They cannot see sail trim, or crew position, or the boat you had to keep clear of. Those answers are in the heads of the people who were there, and the review is written to pull them out.

The same targets are turned round and pointed forwards, as a single card on the display that reads the live numbers and says what to do about them right now. Below is that one card at six different moments of a race, not six different displays: the wind angle and the boat speed against what the certificate wants, and a verdict. Point higher. Bear off. Build speed. Sail deeper. Or, occasionally and satisfyingly, hold exactly what you have.

What This Is Not

It is not a product, and it is not going to be one. It is a few hundred commits of enthusiastic weekend code written for one boat and one crew who already know what the numbers mean. There is no onboarding, no configuration screen and no support. The whole thing runs as a static page because that is the least infrastructure that works, and the hardware is a single-board computer in a bag below decks.

Doing it this way is the point. Nobody would have funded a project to calibrate one fifty-year-old boat’s wind vane. As something a group of people who enjoy this could build between races, it was entirely reasonable, and it answered a question we had been arguing about for years.

The Part We Keep Thinking About

We build AI systems for a living, so the thing that stuck with us is not the sailing. It is that the analysis spent most of its effort deciding whether to trust its own inputs, and then said out loud, in writing, where it still could not.

That is the part that generalises to every piece of data work any of us do. Point a capable model at a pile of measurements and it will produce an answer. It will be fluent, well-organised and confident. Whether it is true depends entirely on whether anyone checked that the measurements meant what they appeared to mean, and that check is the boring, unglamorous work that gets skipped first when a deadline is close.

The other half is in the photograph above. Fourteen people on a wet foredeck, hiking out, calling waves, trimming. We can analyse as much as we like, and the analysis can tell us to point two degrees higher. Somebody still has to hold the boat there for twenty minutes with water coming over the rail. The data does not sail the boat.

Come and Talk to Us About It

There are several regattas between now and the 2027 Rolex Big Boat Series, which will be our fifth as the Typon crew, and we intend to use every one of them: better calibration each time, a debrief the same evening rather than weeks later, and hopefully a boat that points a little higher. If you sail and you have opinions about any of the above, or you are trying something similar with your own instruments, we would enjoy the conversation. Tell us what we got wrong. Get in touch and it will find its way to the crew.