I want to fabricate a new, NACA foil-shaped rudder blade for a tandem kayak and decided to try out the Gemini AI tool to see if it could help. As way of background I was a computer science professor for 32 years, followed by 8 years at Microsoft and a couple years consulting for various computer firms including Facebook, Splunk, Snowflake, and now Amazon. Until today I was truly a skeptic of the entire LLM/AI craze. After a few iterations I used the following prompt with Gemini: “Please design a NACA foil shaped rudder blade for a kayak that is 55mm wide at the top, 570mm long with a straight leading edge, 110mm at the widest point which should be 500mm from the top, and 6mm thick at the widest point”. Gemini produce a 7page PDF document including the NACA CSV profiles for the top, at 500mm and at the tip. Also included were instructions on how to import the model in Rhino and Autocad and fabrication tips. I no longer a skeptic!
Hi Dave,
it is impressive the leaps being made in AI and the reach of even un-specialized LLM’s into areas that we consider ‘specialized’. i have been asking chatgpt (and Gemini and Claude and other LLMs) since the first launch of publicly available chatbots in November 2022 about kayak building and design. and the capabilities today vs three years ago are astounding.
when not kayak building, i am the Chief Information Officer and Chief AI Officer of a federal agency…and candidly, the ability to responsibly use LLM technology to turbocharge your productivity professionally is going to be a basic skill.
my latest project, which i am starting - a new kayak, i recently finished my design selection/design modification phase…and i used LLMs extensively to vet design options against the requirements i had and to explore different approaches in ways that would have required hours and hours of mundane fact gathering and organizing for comparison…
i am looking forward to having moved out of the exploration phase to now moving onto organizng and getting ready for the build 
Please share more! Much more interesting than a new rudder. Can one feed an LLM the cross section offsets and have it produce a 3D model of the hull? If so, can it then run a hydrodynamics analysis of the hull? What about calculating the LWL and capacity?
https://www.theguardian.com/us-news/ng-interactive/2026/jan/18/tech-ai-bubble-burst-reverse-centaur
Laszlo
Redfin Realty has a feature on their website where users can use an AI design agent to operate on a photograph to redecorate a room in any number of styles. Here’s the result of handing it a picture of an actual conventional bathroom with a tub/shower, toilet and a window:
What’s that AI smoking? Aside from the pukey colors, I’m most impressed by the practicality of the arrangement. Base cabinet units stacked on each other at the end of the tub will be convenient for accessing supplies while showering. The hanging light should last about a week in the steam. Any towels on that towel bar should be a mildewed pile in just a couple of days. But my favorite is the pink prison toilet/sink combination with the door for storage in the toilet tank. Great place to keep the spare toilet paper.
This is an AI app on a commercial website meant for everyday use by people making decisions that involve multi-hundreds of thousands of dollars. It’s not just some nerd hacking in their basement. I’d say that the hallucination problem is still a big one..
David, you might want to double check your fin design.
Laszlo
I never said they had solved the problems that LLMs have with hallucinations! That said the fin looks pretty reasonable (see photos) I took the cross sections that Gemini generated, important them into a CAD tool (I use Shapr3D), lofted the complete fin, exported the STL files, and then 3D printed the fin (in parts because of its length). The second photo shows a cross section of the full fin at about the about 6” above the tip. To me the foil looks reasonable. The first photo shows the two halves of the full fin. My plan is to laminate each half and then epoxy the two halves together. The 3D printed part will remain in the fin. I am currently thinking about a layer of 4oz cloth, followed by 2 layers of 6oz carbon fiber, followed by a final layer of 4oz. But I would welcome layup advice from anyone who wants to chime in.
(In case anyone is wondering the gray fin was printed vertically in sections and the layer lines are not noticeable. The black version was printed flat on the print bed andillustrates the limitation of 3D printing convex surfaces as consumer grade printers print in layers that have a minimum size of about .1mm. This really will not matter in my application because the entire surface will get covered in epoxy and fiberglass/carbon and the printed part will remain in the laid up fin)
Looks good. When I print RC airplane parts it’s in vase mode with vertical sections for exactly that reason.
As far as ideas for the lamination schedule, what kind of cloth are you using for the carbon fiber? Uni, biax or woven? Based on my CF mast building experience, I’d suggest uni with the fibers running spanwise. I’d also go with 1 layer of 6 oz and make the glass layers be 6 oz biax.
I don’t have numbers to justify this, but based on how my masts worked out, a single layer of 6 oz CF will be plenty strong. I’m recommending 6 oz glass because of the likelihood of the fin impacting bottom and needing good abrasion protection.
Again, looks good. Looks as if your AI had better training than Redfin’s.
Laszlo
This is the cross section of an Iain Oughtred Grey Seal. It’s an .STL file and only a half model. I also have station frames and diagonals.
I used AI to create, but the process is more akin to OCR.
- Photograph the lines table and ask GPT (Claude, ChatGPT, Gemini, CoPilot) to transcribe to a .md file. it will just reproduce the table in text. I work with .md (markdown) primarily because it’s pure ASCII text and the GPTs all prefer it to having to parse say a word.docx or excel.xlsx.
- Ask the GPT to use the .md file to produce an .STL using python. The GPT doesn’t actually create the model, but it writes python code to generate the model from the lines table.
- You might want to embed the .STL into a webpage with controls for easier viewing, just ask the GPT to write the .HTML as a standalone file (it will embed the CSS and .JS in the header) and embed the .STL with controls.
- Ask the GPT to add controls allowing you to turn on/off stations, diagonals, lines etc.
Note in the picture at station 2 there is an unfair lump. This is a minor error with IOs table measurements. To be fair to IO the measures are all hand drawn and a boat builder in the process of lofting and fairing the lines probably wouldn’t notice the minor error as they fair with a spline. But it shows up in the .STL.
I think this is a little unfair a comparison , Laszlo. Image generation and LLM text generation are different beasts. The real concern though that I think you are trying to get across is the non determinative nature of generation in both vision and text generation. And in this case I think it’s important to use the Generative AI and determinative tools together with each doing what it does best.
For example when David used Gemini to produce his NACA foil it would have written python and used tooling to develop the .CSV from a known table of profiles, again using tooling to copy the profiles from an online source. And only using it’s generative capability to write the python code.
And, yes it’s always good to check your results. But knowing how something works rather than just, ‘I asked a black box to do a thing’ can go a long way to improving trust in the system as well as support exploration of the system capabilities beyond simply asking it to answer a text question with a text reply.
All very true. My real concern is untrained people willing to blindly accept the computer’s work without knowing how it was arrived at and without checking. This has been a problem as far back as I’ve been associated with computers, but it seems much worse today, probably because of all the hype and marketing and tech companies trying to be the next big thing. In the past it wasn’t as consequential if someone took bad advice from a computer but today they’re insinuated into so much of our critical infrastructure that blind acceptance and misunderstanding of their capabilities and functions can have devastating consequences on a global scale. Especially with government and industry’s desire to replace skilled expensive experienced workers with AI. Doctorow’s article articulates all this.
Your half-model design process and David’s fin design are examples of how to use AI correctly. You two are pairing the real stuff with the artificial and achieving solid dependable results. You’re augmenting expertise, not trying to replace it. My point with that ridiculous bathroom was to show what happens when AI is used inappropriately and why you cannot put blind trust in it. With all the hype being thrown at us by Big Tech, I think the occasional reminder of what misuse can lead to is important.
Laszlo
Neat intellectual exercise, but honestly I would not waste the money on carbon fiber and epoxy for this design until you make some significant modifications.
First, it is very much too long and has too much surface area. If you look at the kayak rudders available on the market, you won’t find anything even close to that big. In comparison the Smartrack tandem blade is only about 16” long (only 11” below the water line) and has a wetted surface area of approximately 33in^2. If you look at double surfskis designed to paddle in big waves, the biggest rudder they use is 9-10” long. Your vastly oversized rudder with be a significant source of drag and also a big time weed magnet.
Secondly, recommend modifying the planform to an elliptical shape like that used on the British Spitfire. That shape is widely used in kayak rudders (but not weedless rudders) because it produces a higher lift to drag ratio than other shapes.
Lastly, the foil shape may be wrong depending upon what type of tandem you have and how/where it will be used. The thin foil with pointed leading edge is low drag when in trail (zero angle of attack) so it is good for applications where you will mostly go straight. The trade off is that that shape is not good for turning. The pointed leading edge causes flow separation at just a few degrees angle of attack which creates a significant drag increase. If you look at the rudders on boats designed for maneuvering, they have a much more rounded leading edge and double the thickness of your foil shape.
My comments may seem overly critical but they are based upon both significant professional and personal experience.
This picture shows eight rudders that I commonly use on my boats. Those on the left are under-hull rudders and on the right are over-stern pop-ups. Left to Right: DK 7” semi-weedless for big waves, DK 4” weedless, DK 3” weedless, Stellar 4” weedless, Smartrack Short, Smartrack Single, Smartrack Tandem, Home made attempt at over stern weedless.
Here are the three Smartrack Blades. Note how much smaller the Tandem blade is than your design. Also note the elliptical planform.
The Smartrack blades use a foil shape much like your design. Note the significant drag increase once it has more than about 7 degrees of deflection. I use this on my Mystery and it is fast going straight, but a huge source of drag when turning.
Here is the foil shape of the 7” DK designed for maneuvering in big waves. It is about 14 mm thick compared to 6 mm in the Smartrack blades. Probably a bit more drag going straight but a powerful turner.
No offense taken and I really appreciate your input. The tandem (a cnckayaks.com Vember tandem) is 19’ long and currently has a Smarttrack tandem rudder on it. Frankly, there is not much about the Smarttrack rudder system that I like. I used the Smarttrack pedals once for about 30 minutes and immediately replaced them with SeaLect TruCourse pedals which work much better for me. (I would love to give my SmartTrack pedals away if anyone wants them).
I don’t like the SmartTrack rudder as it stores vertically so I recently bought a SeaLect Trucourse rudder which stores flat. Unfortunately there is no tandem blade available for that rudder system which brought me to the current project. As for length it is exactly the same length from the top of the pivot pin to the bottom of the blade as the Smarttrack tandem blade. It looks much longer because of the way the TruCourse system works.
Both the SmartTrack and TruCourse rudders are 6mm thick. Pursuant to my initial post I have done a V2 design which narrows the blade significantly but have not printed it yet. I would welcome your advice before I go any further. I also did a test layout of the V1 design to see whether the layup using the material I had on hand would be sufficiently stiff but that is still hardening up.
Hey David, I understand your issue now and why you did what you did.
If you had not already purchased the Trucourse housing, I would suggest you look at the other housings that Smartrack sells. They multiple choices to choose from depending on the geometry of the stern including one called the Hybrid Foil that stows horizontally. The one that CLC sells is the Standard Housing which is too large for the low sterns that my race boats have, so I use either the Race or Ski housing on my builds. Here are some of the other choices but the Hybrid Foil is newer and not listed on this 2019 catalog.
Since you already have the Trucourse housing, I suspect that you will want to continue with your printed rudder blade project. As I indicated above, I suggest that you attempt to replicate both the size and shape of the underwater portion of Smartrack Tandem blade. That blade has an excellent reputation and is widely used on high end tandems like those sold by Stellar. I use it on our 18.5’ Shearwater Double and occasionally on the Mystery when paddling in rough water and don’t have any complaints. Assuming that you have the size correct, the biggest thing to fix is the planform shape. The planform of the Trucourse blade is pretty “unique” (not in a good way) so giving it more of an elliptical shape like the Smartrack will be a performance improvement.
Regarding the foil shape, trying to replicate the Smartrack will also be a good place to start. To my eye, the leading edge of your printed rudder is a bit too pointed so I would round it over some. This allow the blade to operate at will a higher angle of attack without stalling. I would also recommend against going any thinner, because a thinner foil will produce less lift. If you plan to paddle much in waves, you may actually want to make the blade a bit thicker. As shown above, the rudders for my race boats used in waves are around 14mm thick.
I hope that this helps!
Thanks. I hope you would you indulge me by answering a few more questions Why is the SmartTrack shape considered so good? Primarily its low drag? Why is the Trucourse blade plan shape considered bad? Too much drag? Not enough lift? The wrong shape for waves? When I look at foils for most high performance small sailboats (e.g. Goat Island Skiff) they all primarily use foils that are much more similar to the TruCourse than the Smarttrack. Frankly, my experience with the Smarttrack is that it doesn’t have nearly enough lift and that more surface area would improve that aspect significantly. I think the blades on the Delta 20T and Passat G3 (both of which I have paddled extensively) are far superior when it comes to control.
If none of this works out I will simply sell the TruCourse housing and blade and purchase a different Smarttrack rudder or just keep using the current one.
The Smartrack rudder blade is efficient because it has an elliptical shape. In the early 1900s they discovered that wings with an elliptical shape produce the least drag. The math behind this is complicated and it takes my Aerodynamics book two pages of equations to explain, so I am not going to try. An elliptical wing has a “span efficiency factor” of 100%. Elliptical wings are hard to build so very few aircraft with that wing shape were built, the most notable of which was the British Spitfire. It is much easier to build a wing with straight edges, so tapered wings came into being. In modern aircraft, they also use aerodynamic and geometric twist to create an elliptical lift distribution like an elliptical wing. Modern wings have a Span Efficiency Facor of 85-95%.
Regarding sailboat/kayak foils, they are much smaller and generally molded so easy enough to build with the most efficient elliptical shape so the are very common. If you look at high performance sailboats like the 420, 470, Melges 22, or any of the Hobie Cats you will see elliptically shaped rudders. The same is true regarding the rudders on racing kayaks, surfskis and outrigger canoes. Those that are not elliptical tend to have some taper towards the end, like the rudder and CB on my Goat Island Skiff. (As a side note, for a home built, the GIS is a good performer but not at all high performance. The rudder and centerboard shapes are easy to build and do not even have a foil cross section.).
Another observation is that I am not the only one who thinks that the Smartrack rudders are efficient, because they are standard equipment on a large number of race kayaks including those built by both Epic and Stellar.
From my perspective, the problem with the Trucourse shape is that the blade gets wider (more chord) as it gets closer to the tip. This puts more lift closer to the blade tip which will strengthen the tip vortices and increase drag. Also note that nobody else uses that shape. If it were good everyone would be using it.
Thanks. I guess since I am not a racer or paddling a surf ski these kind of issues aren’t really all that important to me. All I want is a rudder that will work better than the SmartTrack rudder for me.