0:01 Hello everyone. Okay, my name is Evo, CEO of Braid and today we’re actually going to be presenting together with Guido. Nice to meet you. I’m Guido and the founder. Yeah, essentially the presentation is going to be fairly straightforward. We’re going to talk in the initial stages about what are our main assumptions and the things that we, you know, hold to be truthful. And then for the fun part of all the demonstrations and so on, Guido will come in afterwards.

0:31 Recorded at the CDFAM Computational Design Symposium in Tokyo, October 2026. Learn more at CDFAM.com. And what we’re doing is basically trying to find a reasoning model that can find good solutions for engineering problems. Now I could good. So I think the main idea that we have as a company is always this thought that engineering needs a lot more room for discovery. A lot of time is spent trying to draw a cat, trying to find a good solution on a physics simulator and then picking up our own brains to try to figure out how should I change my problem again and how should I kind of find a better shape by hand and these iterations just keep going.

1:25 So you spend a couple of months and our kind of idea is that the real bit of kind of creativity comes at setting up the problem and finding the good solution in the end. So what we want is to discover these high quality meaningfully different kind of shapes and try to figure out a way of presenting enough information so that you can understand them well. And the way we break it up is again between the definition of the problem, then the automated way of discovering good shapes and then finally the inspection.

2:01 And when you think about AI assisted tools, you typically have human in the loop and you have this idea of going back and forth. We do that for the define and the inspect. But for the discovery, we really try to be machine only without any human intervention. And that’s kind of the key point that we’re going to be trying to cover today. The defining the problem itself is kind of interesting.

2:26 When you think about, you know, problems of design in the real world, for example, a coffee mug, you kind of thought about, I want a vessel that takes 20 cm liters of hot liquid that a human can drink from and put down. And that’s your design problem. You didn’t talk about materials, didn’t talk about the geometry. Didn’t make any choices whether it should be circular or octagonal at the base.

2:51 But on the traditional tools, you’re forced to make those calls almost immediately as soon as you sit in front of a CAD program. We want to kind of change that. Capture the intent as early as possible. Present different solutions as early as possible. And the part of presenting solutions is also very interesting because you want to make trade-offs, comparisons, provide different counterfactuals. Maybe the problem that you set up was too hard and I need to figure out a way of telling you in which direction you should make it easier.

3:21 You know, one thing that happens when clients touch our software for the first time, they set up a easy problem, they get the superhuman response. The next one is going to be an attempt at something absolutely impossible for the physical world. You know, can you design me an airplane that weighs 10 kg and carries 10,000 people? You know, is not going to happen at all. But this is very common that at the point that you’re setting up your problem, you don’t know yet it’s going to be feasible or not.

3:46 So handling those failures and explaining back why it’s a failure and sometimes presenting counterfactuals of if you relax this constraint a tiny bit, here’s an interesting solution for you. That’s kind of the mindset that we have, but we can discuss that a lot more afterwards. The main part that we really care about for today’s presentation because again we’re not going to cannot cover everything is this autonomous bit and this is where it gets a bit different again most AI tools even today are still expected to have a human in the loop validating steering telling you what to do next.

4:21 If you want to actually be able to go through the full design problem, how to figure out how to develop new alternatives, how to evaluate those and based on that evaluation, use the semantics of that evaluation in order to guide the next step. You start to have to think about that level of automation a bit differently. And those tasks are going to be quite long running. And basically today we have sort of like a focus on only five or sorry only three of maybe 20 different ideas that guide us.

4:51 Well, the basic one in the beginning is this notion of very very sparse data. If you think about the amount of cath the like files that you can get access to and then that’s already a pretty small set if you add the problem definition to be paired with that path so that you know that that was the solution for that particular problem. Even smaller set if you add the actual intent that was involved in choosing any of the features you’re getting to an extremely small set.

5:22 So how do you find good ways of you know handling this problem when this is the case and in our minds is to try to mix parts that need to be learned because we don’t have a very good model for and other parts that maybe you can make a satisfactory model for example engineering knowledge and the physics knowledge and the manufacturing aspect from a mathematical perspective I don’t need to necessarily learn again you’ve probably seen these kind of results already on tabletop games even on mathematics recently where you do have a general kind of search mechanism but you still have a validator that is properly formalized.

5:59 Of course for engineering it’s kind of an interesting world because you’re kind of in between the very very formal world of mathematics and the very informal and ambiguous world of natural language. So you can still take some from one side and some from the other side. The other idea too that is interesting for us is that if your substrat is actually expressive enough and you can talk about things at a kind of engineering level, then even general models can help you steer because they can do the same kind of you know decisions that a general human pattern would would make.

6:36 And if you get to the point of saying for example on this particular part add a flange here that’s something that the model can say quite easily. Now can you then verify everything properly with all the guarantees all the manufacturing aspects and all the physics and so on for that particular action. That’s where we put a lot of focus on as well. And if you can do that then you can interpret this kind of decisions that you make and you can interpret the consequences.

7:05 So you get to the point where you can analyze those consequences and keep going on that loop. So the information that you would have given to a human in order for them to act is the same information that you could give to yourself in order to keep acting as well. And the final idea that I’d like to share with you and thank you very much for holding for so long before the cool stuff is basically that if you do have this kind of information, all of your physics that you care about can be treated almost as a compiler.

7:36 All of the mathematics and the manufacturability that you care about can again be treated almost as a compiler. You get to this point where the highle decisions and reasoning system is kind of independent. And when we started was before you know general LLMs and so on. We already found very good solutions beating most human teams. Nowadays you have a better model to guide. We just find them quicker and more efficiently.

8:02 If there’s another reasoning model coming tomorrow we can still plug it in because what we are defining is the rules of the game. And what we built is essentially this kind of framework that allows us to talk about the maths, the physics, the engineering concepts and play around with them so that we can explore these shapes extremely quickly. And again giving this sort of you know experience to the final user.

8:27 Very creative moment for engineers when they define the problem and think what they want to solve. The second very creative moment is when I have all of these different alternatives and I need to figure out what are meaningful tradeoffs, which one am I going to present to the team, which ones that we’re going to keep moving forward with. And the in between is just 2,000 clicks on a CAT program and 10,000 pages of, you know, reading manuals in order to figure out how to use a physics simulator.

8:54 That part can be completely abstracted away from the human. At least that’s our idea. And I’ll let Guido present to you a bit more of what we do. Thank you. Perfect. Thank you. Let’s go for the high candy part of the talk. Yes, first of all, yeah, I’m going to concentrate on this discovery part. So there are a lot of interesting ideas on how you’re going to help the user define problems and how you can help the user like inspect and look at the solution that he found.

9:24 But today we’re going to look at what can you discover when you start applying these ideas to an automated engineering system. Before going to the application ideas it it would be great if there was a literally a magic box. You give me the problem definition something or something happens in the middle I get some options at the end and literally requirements requirements requirements in and manufacturable designs out.

9:50 So that’s that’s what we are working on at braid and let’s open the box. This is this we’re calling this box aura at the moment and there are two major layers and it’s nothing particularly new here. If you have to support your physics, you have to support your geometry constraints and so on. You will have to have some multiple geometry generators in your system, multiple physics solvers depending on the physics depending on the way you want to solve the same physics for for example and several ways to search this giant space of possible possible choices that you have and so the point is that someone has to choose which one I’m going to use of these elements and the current way is it’s human in the way in in in in the loop always.

10:41 So the important layer is is a reasoning layer that ingests all your problem definition and decide makes autonomous calls on what I am going to choose and actually going to run and automatically and autonomously. So the reasoning part is like literally is the same as what what choices I’m going to make and what is the reason why I’m making these choices and that has to be automated fully automated if you really want to achieve this like magic box kind of solution.

11:15 Okay. So this was a very brief introduction about how we think about hor it’s very very high level and in terms of how it looks in practice in terms of like what are the kind of results that you can actually discover. I’m going to focus on just three three cases on this talk assemblies. So literally talking about can we generate multiple parts at the same time and this part each part has interdependency with with each other and can we handle complex requirements specifically a manufacturing requirements like mass manufacturing requirements and complex physics like nonlinear physics and and multiple additional constraints and and what and can we handle like the m the manufacturing methods that are relevant for the the large part of the industry.

12:07 Okay, let’s start with the assemblies. So this is a case of a a door panel. The the idea here is that you have some fixed some fixed invariant regions. And what you want to design is that the actual panel and your your anti-intrusion beams and there could be more than one anti-intrusion beams. And all of those needs to be designed with respect to crash analysis and additional like stiffness they could be in contrast with your crash analysis and so on.

12:43 So all of these like multiple constraints and the important part is these parts will be interdependently will be the design of each part is dependent on the on the other parts. So and if you run our system, this is the kind of designs you’re going to get. So you have your panel and and one or multiple beams that are generated towards the panel too and and and vice versa.

And and this satisfies all the requirements on on physics and manufacturability and and as I was u well yes so this is the first case in terms of the handling assemblies. So another interesting case is can we handle extremely complex constraints in terms of manufacturability and physics. In this case this is a this is a hood design. The the idea here is that you are given by your stallion team the outer part of the hood.

13:45 You can change that and you need this design internal part and the internal part needs to be stamped in this case steel has to satisfy a lot of different constraints in term of how can I what what my my machine can actually do and and there there needs to be design like the bonding between the two parts and all these manufacturability constraints that are required and and on the physics the the hardest constraint is the head impact you you don’t want a a pedestrian when when there’s a there’s a collision with the pedestrian, you don’t want the pedestrian to be badly hurt.

14:17 And there is actually in contra in contrast with making this thing stiff. So there are multiple competing constraints and extremely nonlinear physics that needs to be handled all at the same time. So given this is the brown definition. Yes. So the brief video. So you have your stallion team that gave you this. What we need to design is the internal part and and this part needs to be designed together with all the mastics and all the constraints on your on your manufacturability and specifically draft angles elongation like how much you can press this this this part.

22:20 And once you have all of these together and specifically you have to place your mastics dependent on the design you can start generating design and the the part is that the geometry generator will by construction include all these constraints. So what you’re going to generate is manufacturable by design and you can generate multiple of those but the the important aspect that also all the physics needs to be need to be constraining the final choice of your of your design and that’s that’s where the crash analysis comes in that this information needs to be fed back to your search algorithm in order to actually figure out what you can do and of course this is completely independent of what is going to be the initial outer design that your styling team is giving you.

And if your styling team decides to change a little bit the design, well, you can just rerun this and in a couple of days you actually get the new set of woods that you can validate at the end. Okay. So moving ahead from the all these complex requirements I want to focus on this like fast last part on several manufacturability constraints on your geometry that we are able to include in our in our reasoning agent.

So this specific is is stamping and this is the same part multiple stamping that flanges beads and and of course the satisfy physics in in in the in these cases like bending and torsion. It’s not a very complicated problem but it’s just to showcase that we can handle like stamping requirements and even harder requirements is sheet metal bending. Sheet metal banding is you’re forced to have bands only on a on which are linear on a straight line and all the rest has to be a planner.

So this is a a more complex constraint than the than the simple stamping and and this is the kind of geometries we can generate and given your problem definition you give me all the areas which are in gray here it’s very simplified kind of model here all the areas that you need avoid and of course I can provide you like the bracket that satisfies the physics satisfy the manufacturability and so on and in terms of the manufacturability taking one of those you need to satisfy the fact that literally can has to be produced in a specific sequence and and this has to be taken into account when you actually generate your part.

And another interesting part here is that the system realized that I need to create this band here you can see yes I need to create this band here. So I need to create this notch which going to be essentially a band relief during the manufacturing the manufacturing process and this is very important and specifically also created this like lamp of mass here and this was literally created just because there was a vibration requirement and this this increased the actual frequency so you can get out of resonances continuing on this.

So this the idea here is that you kind of have the same problem and you want to provide the user as Evo was mentioning a way to discover the your space of possibilities. So even when the user doesn’t specify that oh I want this specific manufacturing method we can provide an an exploration of of severals that are close enough then they could be they could be relevant for the specific problem.

And in this case the same bracket the well the same connector has been solved in in at least two different two different manufacturing methods which is bending and press forming and the solutions are extremely different. There’s you have the lightest one and then you have like some others which are pretty weird but this is like just 4 mm for example. And actually the interesting aspect of of this specific problem was that the exploration was going to do an exploration on several thicknesses but for example up to 10 mm there was the exploration and this specific agent decided to not use 10 mm because a buckling analysis showed showed that it was not possible to actually even manufacture.

So that was completely excluded by the by the by the by the system. So there are multiple multiple considerations on manufacturability that all taken into account before providing the final few results to the user. You don’t want to provide we we since and going connecting back to the to the case of of discover discovery and discoverability. We want to provide the user a a a panoramic view of what is possible given the problem definition even when the user didn’t fully specify the the problem and which is most of the time the problem is always slightly underspecified.

Finally in a completely different case this is like a simplified case very simplified case of of injection molding. In this case, we just creating the rib pattern below and and but we are the the system is is fully able to generate these kind of structures which are manufacturable by construction and in this case is also generating the the two dies. Again it’s a very simplified problem different from the what what we were working towards with the client.

Yeah, summarizing essentially so we I presented a few cases on how our system discover new shapes and what kind of shapes can be discovered in this case and but there is as I was mentioning before there is a lot of work very interesting work on the define and inspect specifically on the UI if you really want to be helpful for the for the user even during the definition of the problem and not being like the user driven by the program but the user driving the program and driving the the software in terms of this is the problem I’m defining and then the the the the UI actually changes as soon as I’m defining the problem we didn’t discuss this today there will probably be another another talk by itself and same on the inspect once you have discovered a lot of shapes you like to and you discovered and you had a a a reasoning that helped you discover and this was the automated part and you want to reuse that information to actually provide relevant relevant feedback to the user.

Okay, this part has had this reason because it was it was designed this way because of this reason. If you can provide that information, that’s extremely useful for the engineer to actually think about the problem and eventually add more variation on the on the definition or go to the next step and so on or even just learning from the fact that the machine decided to generate that specific part.

Finally, yes. So, manufacturing will define what has can be generated. And if you’re talking about mass manufacturing, that is extremely strong constraint on on the geometries you you’re going to be able to actually generate. And once you have your geometry generators that are respecting manufacturing constraint, all what the physics and additional requirements do, they steer this kind of discovery and search in this like insane space of possible shapes.

22:34 And another important part is that you like to start since we are we don’t want to build a system that is like retrieving information from some memory and so we want to start and want to give the ability to generate completely new part on that was not taught even before. So it’s not like reusing engineering information that’s been accumulated for 20 years. Maybe we can actually discover new things.

23:04 You really need to start from first first principles. So your system needs to be working from like very basic algorithms so you can control and that’s where the trust on the final design comes to. Yeah. So first of all, thank you very much for listening and for coming here and we hope that the rest of the conference is going to be great. Thank you very much. Recorded at the CDFAM computational design symposium in Tokyo October 2026. Learn more at CDFAM.com.