Bobby Ng STEPs Up with Transfigure

WEBVTT
00:00:07.360 --> 00:00:11.919
Hello and welcome to FoDES, the Future of Design and Engineering Software podcast.
00:00:11.919 --> 00:00:13.679
My name is Roopinder Tara.
00:00:13.679 --> 00:00:20.480
On the show, we will have guests that will discuss tools and technology that engineers will find interesting and useful.
00:00:20.480 --> 00:00:22.480
Bobby.
00:00:22.480 --> 00:00:23.600
How are you?
00:00:23.600 --> 00:00:24.719
I'm very good.
00:00:24.719 --> 00:00:25.760
Nice to see you.
00:00:26.079 --> 00:00:28.719
Yes, I'm super excited for our chat.
00:00:28.719 --> 00:00:30.239
I've been thinking about this all week.
00:00:30.480 --> 00:00:32.000
Ah, welcome to the show.
00:00:32.479 --> 00:00:38.000
I just have to say, like I'm more nervous than anything I've, you know, I pitch and I have to sell and all the stuff.
00:00:38.000 --> 00:00:46.399
I'm actually more nervous for our chat than anything else, but I'm also the most excited to talk with you because um you have some great content on YouTube.
00:00:46.399 --> 00:00:51.920
Really appeals to certain type like myself and CAD junkies.
00:00:52.000 --> 00:00:53.679
So oh well thank you.
00:00:53.679 --> 00:00:56.320
Thank you very much for thanks for being a fan.
00:00:56.320 --> 00:00:58.479
Yeah, uh, really appreciate that.
00:00:58.479 --> 00:01:03.039
Uh, but really, I'm just, you know, you guys are doing all the smart work.
00:01:03.039 --> 00:01:10.319
I'm just I'm just watching and recording and every now and then offering a pearl of wisdom, if I may call it that, even.
00:01:10.879 --> 00:01:26.640
I think it I think it was a lot of wisdom there because honestly, I you know, I think the reason why I feel nervous is because you you've seen a lot, um, more than probably most VCs, most other founders or technical people get to see of this world that we're kind of gonna dive into.
00:01:26.640 --> 00:01:29.200
So I feel like you have the most context.
00:01:29.200 --> 00:01:32.719
Like what's good, what's bad, who's cool, who's not cool.
00:01:32.959 --> 00:01:35.439
So I'm trying to figure it out best I can.
00:01:35.439 --> 00:01:46.879
And uh I'm not saying uh we're doing a great job because tell you the truth, there's just a lot to cover, and I feel like I'm able to get to the very the surface level of it.
00:01:47.519 --> 00:01:58.799
We're working in a very different area, which I think you you and your the audience will find out about pretty soon, but um just the whole world is changing even faster than um any other time in the past.
00:01:59.040 --> 00:01:59.760
Yeah, it is.
00:01:59.760 --> 00:02:09.199
I mean, just uh November 30th, 2022, GPT was announced, and then the world I can like to say the world like split in two.
00:02:09.199 --> 00:02:15.520
They used to say before BC and A D, and now it's like before AI and after AI.
00:02:15.520 --> 00:02:16.879
Yeah, something like that.
00:02:16.879 --> 00:02:17.360
Yeah.
00:02:17.360 --> 00:02:26.080
So we should we should go into this in more depth because a lot of the stuff I've been hearing about, everybody says okay, LLMs are no good for CAD.
00:02:26.080 --> 00:02:32.960
Uh they're just about language, and then there's a whole bunch of whole bunch of AI startups that came out.
00:02:32.960 --> 00:02:39.360
Michael Finocchiaro, I think, does the best job of logging them and taking notes of them and talking to them.
00:02:39.360 --> 00:02:44.159
And he's recorded over, I think what's the number now, 880, I think.
00:02:44.159 --> 00:02:49.280
AI startups in our industry, our industry may be very broad, CAD, PLM.
00:02:49.360 --> 00:02:49.840
Yeah, right.
00:02:50.080 --> 00:02:53.039
Even yeah, I think he goes even further, manufacturing.
00:02:53.039 --> 00:02:54.000
Right.
00:02:54.000 --> 00:03:03.439
I gotta say, most of them, most of the startups that are in this CAD field are devoting themselves to text to CAD.
00:03:03.439 --> 00:03:10.800
And you said very clearly to me, no, we're not that, and we're not, we're not all labs, and we're not, is that correct?
00:03:10.879 --> 00:03:38.319
You're not and and I I initially misinterpreted what you said to me because you said something like you like how we're not using AI, but I think you put not in capital letters, and I misread that and I said, Oh, but we are using AI, but then I realized later, I'm like, oh, we are not using it in the way other people are, and in fact, we're actually doing something that is totally unprecedented in the world of CAD, uh, which I will I will do my best to explain.
00:03:38.400 --> 00:03:41.759
Uh I did go to the into the XF.
00:03:41.759 --> 00:03:43.840
I tried gave it a shot.
00:03:43.840 --> 00:03:44.400
Yeah.
00:03:44.400 --> 00:03:48.319
So and I'll I'll comment on on what I what I thought of it.
00:03:48.319 --> 00:03:49.039
Yeah.
00:03:49.039 --> 00:03:50.400
Okay.
00:03:50.800 --> 00:03:55.919
So um, how would you like us to go through the next, I don't know, 30 minutes or so, Roopinder?
00:03:56.400 --> 00:04:03.039
Oh, give me a give me a why don't you tell us a little bit about yourself and what led you to to make Transfigure?
00:04:03.360 --> 00:04:03.759
Sure.
00:04:03.759 --> 00:04:12.560
So about now it was over a year ago, you know, a co-worker and I, at my former employer, we were having a conversation.
00:04:12.560 --> 00:04:15.039
He's a fabricator, machinist, model builder.
00:04:15.039 --> 00:04:18.480
So he's really he's running the mill, he's welding, he's cutting.
00:04:18.480 --> 00:04:24.079
I'm the designer, and I give him all my stuff, and we said, Hey, you know, AI is supposed to take our jobs.
00:04:24.079 --> 00:04:31.279
That's what they say, except unless you're a plumber who's building AI infrastructure, or an electrician who's wiring AI infrastructure.
00:04:31.279 --> 00:04:34.319
Hey, let's see if let's see if we're in trouble here.
00:04:34.319 --> 00:04:37.120
And so we went to uh our favorite LLM.
00:04:37.120 --> 00:04:39.519
I took Gemini because I just happened to pay for it.
00:04:39.519 --> 00:04:40.639
He used ChatGPT.
00:04:41.279 --> 00:04:55.040
And keep in mind, this is a just is just over a year ago, and the prompt was to make a cube out of as a STEP file, you know, standard format, make a cube one inch square and send it back to me so I could open it in CAD.
00:04:55.040 --> 00:05:03.759
And it's churning and is doing its thing, but and it it tried to directly write a STEP file out token by token.
00:05:03.759 --> 00:05:07.680
And it said, here's a file, you know, it should open perfectly, let me know if there's any problems.
00:05:07.680 --> 00:05:11.199
And you open it, and I'm a KateA user, you know, he was using solve.
00:05:11.199 --> 00:05:13.120
Um open it up and there's nothing in there.
00:05:13.120 --> 00:05:18.800
It's like it it opens, there's a header and a footer, but the data that's inside is no good.
00:05:18.800 --> 00:05:23.839
So reprompt the model, saying, like, there's nothing in there, fix it and send it back to me.
00:05:23.839 --> 00:05:25.279
It's turning, it's doing its thing.
00:05:25.279 --> 00:05:26.480
It's like, oh, you're so right.
00:05:26.480 --> 00:05:28.800
I totally messed up um the file.
00:05:28.800 --> 00:05:29.600
Here it is again.
00:05:29.600 --> 00:05:30.800
It should work perfectly.
00:05:30.800 --> 00:05:34.879
Let me know if there's any problems and open it up again, and there's just nothing in there.
00:05:34.879 --> 00:05:38.959
I said, okay, we're safe, we're good for for now.
00:05:38.959 --> 00:05:45.920
And I forgot about it for the rest of the day, but then later I went home and I was like, What it looks so close to a STEP file.
00:05:46.720 --> 00:05:47.920
Why did it not open?
00:05:47.920 --> 00:05:49.680
And looked at it very carefully.
00:05:49.680 --> 00:05:58.639
And as you probably know, and most of the audience knows, you know, you could read STEP file line by line and try to parse it out and reconstruct this in your mind.
00:05:58.639 --> 00:06:01.360
And but then looking at it, it's clearly broken.
00:06:01.360 --> 00:06:10.160
You know, STEP files are a graph structure format, it is not a token by token prediction, it's a generated file, right?
00:06:10.160 --> 00:06:14.399
So I said, well, it actually knows kind of what it should look like, but it's totally wrong.
00:06:14.399 --> 00:06:19.839
Like it none of these nodes line up and none of the attributes really line up the way it should.
00:06:19.839 --> 00:06:22.879
And I decided, well, maybe I need to train my own model.
00:06:22.879 --> 00:06:28.240
Uh and we took an LLM approach, just train an LLM to write STEP very well.
00:06:28.240 --> 00:06:29.519
And I did that.
00:06:29.519 --> 00:06:45.040
I, you know, I with the power of AI, it's rather accessible now to make a tiny model on your own and train something that can ingest, like input a 2D drawing of a cuboid in certain proportions and output a STEP file in similar proportions.
00:06:45.040 --> 00:06:48.160
And that pipeline totally worked, but it's just cubes.
00:06:48.160 --> 00:06:57.120
But it did prove a point that there is a world in which you know machine learning has a capability of understanding these files, but we need to go way deeper.
00:06:57.120 --> 00:07:06.800
And so LLMs are just not the final approach, like that's one way to do it, but LMs can't write these very long range dependency graphs.
00:07:06.800 --> 00:07:15.519
You know, as you know, there's you know, one STEP file might be a few megabytes, a whole car would be like three or four gigabytes, depending on how much stuff is in there.
00:07:15.519 --> 00:07:25.600
So the ability for an LM to really understand that full world is is fundamentally like it is isn't maybe not unlimited, but is terribly inefficient.
00:07:25.600 --> 00:07:32.000
And so that's why we're taking a very different approach to what we're doing with CAD, which is partially on display at XFGR.ai.
00:07:32.000 --> 00:07:39.199
But we're going a lot further now with a foundation model that we're training that is bespoke to us, very different approach from LLMs altogether.
00:07:39.199 --> 00:07:40.399
So that's what we do.
00:07:40.399 --> 00:07:42.959
We take 2D drawings and we turn them into 3D STEP files.
00:07:43.120 --> 00:07:43.439
Yeah.
00:07:43.600 --> 00:07:43.920
Yeah.
00:07:44.160 --> 00:07:44.480
Okay.
00:07:44.480 --> 00:07:46.240
Got a whole so many questions.
00:07:46.240 --> 00:07:48.319
And we have very little time.
00:07:48.319 --> 00:07:51.920
So I could take up all your whole day, but I won't, don't worry.
00:07:51.920 --> 00:07:58.399
But all right, so is it not text to to CAD or is it like drawing to CAD or sketch to CAD?
00:07:58.399 --> 00:07:59.360
What would you call it?
00:07:59.600 --> 00:08:01.839
This is 2D to CAD.
00:08:01.839 --> 00:08:04.160
So it doesn't have to be just a drawing.
00:08:04.160 --> 00:08:18.800
Uh, it does not have to be uh it could be a sketch, it could be a photo, it could be an engineering drawing, which is great because those are are all within proportion, black and white, and like a nut is a nut, a bolt is a bolt.
00:08:18.800 --> 00:08:39.200
But we'll it will take all of that because just imagine a world in which an engineer were to give you, or you would give a fabricator a hand drawing of you know a sheet metal fabricated bracket, and humans by nature can't draw perfectly straight lines, and we'll just draw something that resembles the part, but really communicating an idea.
00:08:39.200 --> 00:08:52.799
And then if you give that to a CAD drafter or a fabricator, they'll understand, okay, I understand this is a a bracket that's a hat bracket, and it's a hundred millimeters on each side, and the phalanges are 20 millimeters.
00:08:52.799 --> 00:08:54.320
Okay, I'll go CAD it up.
00:08:54.320 --> 00:09:27.279
But that implicit understanding of okay, I understand this 3D object in my brain, which is like the neural net, and I can go to a CAD program and spend you know a few hours doing all the operation serially, you know, pad, uh fill it, extrude, and just so on and so on, and to get this file, those things are human nature, but the way that computers want the way that computers could construct 3D files is fundamentally slowed down because humans have to interact with a CAD kernel in the middle of it all.
00:09:27.840 --> 00:09:35.519
I would call it not I would call it uh not human, but engineering knowledge, right?
00:09:35.519 --> 00:09:37.200
So it's like yeah, you have that.
00:09:37.200 --> 00:09:39.679
You're a mechanical engineer by education, right?
00:09:39.759 --> 00:09:40.080
That's right.
00:09:40.320 --> 00:09:51.440
So you you understand that if it's sheet metal, you know, you you work in units of one bend thickness or you know, one thickness band or two one and a half thickness band or something, right?
00:09:51.440 --> 00:09:55.759
You know that, which by the way, your software was not able to do.
00:09:55.759 --> 00:09:58.879
You know, we are I'll tell you something.
00:09:58.879 --> 00:10:04.960
Okay, I'm getting to this faster than I intend to do, but you we've you brought it up actually, right?
00:10:04.960 --> 00:10:13.440
All right, so it's your own example, and it's the it's a bracket, it's a bracket, it goes like this, right?
00:10:14.399 --> 00:10:15.039
Z bend, yeah.
00:10:16.559 --> 00:10:19.360
It doesn't hold a one bend or even a one and a half bend.
00:10:19.360 --> 00:10:21.279
It does not, it does not, right?
00:10:21.759 --> 00:10:23.039
It does not, you're correct.
00:10:23.039 --> 00:10:25.840
And you know, so this is just our MVP.
00:10:25.840 --> 00:10:32.080
Um we have deployed it so that people can try to use it and just see what's going on a little bit under the hood.
00:10:32.080 --> 00:10:34.159
But you're correct, we don't quantize the output.
00:10:34.159 --> 00:10:44.080
But you're 100% correct, though, that in the engineering world and the manufacturing world of things, there is not an infinite number of tools of sheet metal upper and lower dies.
00:10:44.080 --> 00:10:53.840
There's not an infinite number of thicknesses of sheet metal that are standardized, it's 10, 11, 14 gauge or 1, 2, 1, 1.2, 1.4, and so on.
00:10:53.840 --> 00:11:03.840
So in reality, the way that we would design things for mass production actually is a much more finite space than I think maybe most people would appreciate.
00:11:03.840 --> 00:11:05.840
Yeah, because we have to do it that way.
00:11:05.840 --> 00:11:10.480
Otherwise, nothing in the world would be economically viable if we made everything custom all the time.
00:11:10.480 --> 00:11:23.200
So that pipeline is bridging the gap between artist artistry, trying to preserve an idea versus a really like something that you would put into manufacturing and expect it to come out at a reasonable cost.
00:11:23.200 --> 00:11:25.840
So, because there's another example in there, which is the cow.
00:11:25.840 --> 00:11:27.679
I mean, you saw I saw that.
00:11:27.679 --> 00:11:38.000
You know, we have to really weigh the difference between like how much do we try to make this a geometrical cow versus try to preserve the artistic intent behind it.
00:11:38.000 --> 00:11:41.519
And that is a non-trivial task, I'll just say that.
00:11:41.679 --> 00:11:42.080
Right, right.
00:11:42.080 --> 00:11:50.240
So, okay, so I'll try to make this brief because I've had this request made of me for uh 30 years.
00:11:50.240 --> 00:11:53.759
I've been in the CAD for more longer than that, actually.
00:11:53.759 --> 00:11:56.720
And okay, so here's what happened.
00:11:56.720 --> 00:12:01.919
I was at Autodesk University and I thought, oh, I'm a I'm uh I'm hot shit, right?
00:12:01.919 --> 00:12:06.879
I know enough to get into Autodesk University and you know teach a course there, right?
00:12:06.879 --> 00:12:12.639
And on the way there, a guy going up the elevator with me says, Oh, you're in the you're in the CAD.
00:12:12.639 --> 00:12:16.320
Okay, and by then the world didn't even know how to spell CAD, right?
00:12:16.320 --> 00:12:20.480
But essentially he said, Oh, can I give you a sketch?
00:12:20.480 --> 00:12:26.960
And it you can make me a, you know, the CAD will make it into a floor plan into a into a house.
00:12:26.960 --> 00:12:30.399
I got and again, no, you don't know what you're talking about, right?
00:12:30.399 --> 00:12:34.159
And but I realized that his need was that basic.
00:12:34.159 --> 00:12:41.120
CAD was supposed to be understand what you're intending to do with this sketch, right?
00:12:41.120 --> 00:12:44.480
He could draw out, he could draw out a floor plan, right?
00:12:44.480 --> 00:12:48.639
But we expected CAD to be gonna do a model, right?
00:12:48.639 --> 00:12:52.320
Give me, or at least give me floor plan and elevations, right?
00:12:52.320 --> 00:12:56.080
At least make the walls like standard size thickness, and right.
00:12:56.080 --> 00:12:58.799
He said he expected CAD to do all that, right?
00:12:58.799 --> 00:13:02.159
In 30 years, it never went to that, right?
00:13:02.159 --> 00:13:03.679
It never went to that.
00:13:03.679 --> 00:13:11.360
I cannot to this day give it a front, top, and right side view of a part and have it make a part, right?
00:13:11.360 --> 00:13:16.639
Which is what the world out there, the people, the guy on the street, guy in the elevator.
00:13:16.639 --> 00:13:19.440
That's what he expects you to CAD to do, right?
00:13:19.440 --> 00:13:23.519
It's supposed to be computer-aided design, right?
00:13:23.519 --> 00:13:27.840
I think now with AI, it could actually be smarter, right?
00:13:27.840 --> 00:13:29.039
It could be, right?
00:13:29.039 --> 00:13:30.799
It could be that's what you're trying to do.
00:13:31.039 --> 00:13:33.759
Yes, and we're trying to get there not just for OEMs.
00:13:33.759 --> 00:13:41.519
Like I worked at Honda and Tesla and so on for for many years, like over 20 years I've spent in automotive and a little bit of airplane stuff.
00:13:41.519 --> 00:13:48.480
But you're absolutely right that you know, when when I wouldn't say I went to school to become just a CAD operator, yeah.
00:13:48.480 --> 00:13:49.679
I don't think most people do.
00:13:49.679 --> 00:14:01.519
Some people do, and they're that's their profession, but my job is to solve problems as an engineer, or you know, like my wife's job is to she chose she chooses to try to take care of the house.
00:14:01.519 --> 00:14:04.480
And you know, she sometimes needs things fixed.
00:14:04.480 --> 00:14:14.000
She's like, Can you make another wall plate on your 3D printer to like because one of the kids broke it, you know, smashed a toy into the wall plate and now there's electricity coming out.
00:14:14.000 --> 00:14:16.720
Yes, I could absolutely do that, but I actually don't have time.
00:14:16.720 --> 00:14:31.360
So there's a whole thing of like not enough time in the day, not enough skills, because it does take uh 10, five years to master a skill, 10,000 hours, like to get really good at Katia, you have to put in the time.
00:14:31.360 --> 00:14:36.480
Or the license is just way too expensive, which is definitely the case for for NX or Katiya.
00:14:36.480 --> 00:14:40.240
You know, those are upwards of 35 grand uh per pop.
00:14:40.240 --> 00:14:46.879
So many people get stuck in this situation of they can absolutely draw, they could definitely take a photo.
00:14:46.879 --> 00:14:49.120
I think most people take a photo with their phone.
00:14:49.120 --> 00:14:57.840
And now for AI models to be able to segment those drawings out and create that in a 3D latent representation, those models are there.
00:14:57.840 --> 00:15:11.039
But the final gap, and which is what we're working on right now, is getting from that latent space representation, which is like really the the features that are engineered in the the layers of the the neural net to a STEP file.
00:15:11.039 --> 00:15:18.559
Um no one has bridged that chasm yet, and we're going to do it with our proof of concept foundation model, which we're working on now.
00:15:18.559 --> 00:15:31.360
So we're almost we're getting there, we're just starting on this journey, but it it will take a little time to yield fruit from what is essentially an unprecedented AI model that actually makes STEP files directly.
00:15:31.759 --> 00:15:40.720
Yeah, so okay, so your goal is to make STEP files, read STEP files, and uh that way it could work with be imported into any any CAD system, right?
00:15:40.720 --> 00:15:41.360
That's right.
00:15:41.360 --> 00:15:45.279
Yeah, and we feel that it would be still wouldn't be parametric though, right?
00:15:45.279 --> 00:15:47.120
Is that the next STEP, maybe?
00:15:47.440 --> 00:15:51.120
You know, that I have seen some people say that STEP files are parametric.
00:15:51.120 --> 00:16:05.200
I mean, I guess you could consider it that if you say modifying a graph is somehow parametric, but you know, fundamentally, I think I think people think of parametric stuff as being what PTC originally named their company, right?
00:16:05.200 --> 00:16:13.679
Like, and I learned on auto, I learned on Pro E um a long time ago back in college, before they went through all the name changes and now they're something else.
00:16:13.679 --> 00:16:17.759
Um but parametric CAD is what most of people in my generation grew up with.
00:16:17.759 --> 00:16:36.320
We expect it to work that way, but it is just a means to an end because most people don't have the time to learn how to make a feature tree or you know spec tree for CAD, but they absolutely know how to draw a part and getting the STEP file at the end, which is really what everyone's really trying to create, the actual shape of the part.
00:16:36.320 --> 00:16:43.919
Then you can go to your 3D printer at home, you could send it to Xometry and get it in the mail, or you could actually go mass produce the part in a factory.
00:16:43.919 --> 00:16:51.360
And um, you know, what people do with STEP files beyond that is really up to them if they really do want to put it back into SOLIDWORKS and do the feature recognition.
00:16:51.360 --> 00:16:57.039
Or NX has some other thing called like find feature or something.
00:16:57.039 --> 00:16:59.039
KTA has their own flavor of that.
00:16:59.039 --> 00:17:07.599
Few fusion has a thing, everyone has their own sort of like let's parameterize the STEP, and generally they come out okay enough.
00:17:07.599 --> 00:17:12.640
Or if you're just a clever engineer, you you know how to modify STEP files if you really have to.
00:17:12.640 --> 00:17:22.319
It's not convenient, but um, getting 90% of the work done is oftentimes a huge time saver in and of itself for any anybody.
00:17:22.400 --> 00:17:29.440
So the best the best one I've seen the recognized features was uh actually solid edge with their synchronous technology.
00:17:29.680 --> 00:17:30.319
Okay, right?
00:17:30.480 --> 00:17:39.359
It could go into dumb solid and just you know, and start you could start pulling uh fillets down to to size or move it, move the surface, right?
00:17:39.359 --> 00:17:43.440
And recognize, oh, this is a flange, it's got to move this way, that kind of stuff.
00:17:43.440 --> 00:17:43.920
Right?
00:17:43.920 --> 00:17:45.200
It's it's pretty smart.
00:17:45.200 --> 00:17:46.079
It's pretty smart.
00:17:46.079 --> 00:17:53.920
It's still I'm still, I'll tell you, I'm still frustrated with all CAD systems because because I don't know anymore.
00:17:53.920 --> 00:18:01.359
I don't know, I don't use them enough to retain my mastery of them if I if I ever did, you know.
00:18:01.359 --> 00:18:03.039
But I just don't have that.
00:18:03.039 --> 00:18:06.240
So each time I've got to go learn CAD language, right?
00:18:06.240 --> 00:18:07.359
I gotta learn CAD.
00:18:07.359 --> 00:18:11.599
So my my bone to pick, and I have many, right?
00:18:11.599 --> 00:18:14.400
It's like, why can't I just tell it what to do?
00:18:14.400 --> 00:18:14.799
Right.
00:18:14.799 --> 00:18:20.079
I I want to use a natural language interface in in all my CAD programs, right?
00:18:20.240 --> 00:18:20.559
Yeah.
00:18:20.720 --> 00:18:25.359
So like I want it to say, hey, you know, like okay, I used example, make me a bracket.
00:18:25.359 --> 00:18:28.000
I have issued a challenge, make me a bicycle, right?
00:18:28.000 --> 00:18:32.319
Or modify this bicycle to where it's optimum, right?
00:18:32.319 --> 00:18:37.680
It uses the least number of tubes, and and by the way, nobody's taken me up on that challenge, right?
00:18:37.680 --> 00:18:42.799
Officially, a lot of companies have said, okay, I that's an easy problem to fix, right?
00:18:42.799 --> 00:18:44.960
I can optimize your bicycle frame.
00:18:45.200 --> 00:18:46.880
It would make you a unicycle, I think.
00:18:46.880 --> 00:18:49.519
It would make you a unicycle with a training wheel, probably.
00:18:50.240 --> 00:18:52.000
Maybe that's funny you say that.
00:18:52.000 --> 00:18:53.920
I just saw one go down the street the other day.
00:18:53.920 --> 00:18:55.119
Can you believe it?
00:18:55.119 --> 00:19:02.480
I just saw a guy who actually unicycle, the big one, the pet, not a no, sorry, not a unicycle, but a penny farthing.
00:19:04.160 --> 00:19:05.920
The people that ride those are very special.
00:19:05.920 --> 00:19:08.400
I've known one person and he was a special guy.
00:19:08.400 --> 00:19:09.039
He had a penny.
00:19:09.680 --> 00:19:10.400
Are you a cyclist?
00:19:10.400 --> 00:19:11.599
You sound like you know that.
00:19:11.920 --> 00:19:24.079
I used to be in I used to be into downhill mountain biking of and other types of cycling, but um after a a big missing a big jump and then almost shattering my heels, my wife told me it is time to stop.
00:19:24.079 --> 00:19:25.119
So I stopped.
00:19:25.119 --> 00:19:28.319
But I have kids that are maybe up that age almost now, so we'll see.
00:19:28.559 --> 00:19:29.200
You had to give it up.
00:19:29.200 --> 00:19:30.640
You're a crazy downhiller.
00:19:30.640 --> 00:19:31.680
Wait, what did you break?
00:19:31.680 --> 00:19:32.240
Did you say?
00:19:32.480 --> 00:19:33.839
I almost shattered my heel.
00:19:33.839 --> 00:19:35.359
I got a your heel.
00:19:35.359 --> 00:19:42.400
Yeah, it was you know, the when the bike disappears from beneath you, there's not much else to rely on except your butt or your legs.
00:19:42.799 --> 00:19:43.759
Oh man, okay.
00:19:43.759 --> 00:19:46.319
We could we could definitely this could be a whole topic.
00:19:46.319 --> 00:19:47.680
I've got a collarbone.
00:19:47.680 --> 00:19:49.440
I broke it five times.
00:19:49.440 --> 00:19:52.480
And I'm not even a downhiller, I'm just a road cyclist.
00:19:52.480 --> 00:19:53.119
Oh my gosh.
00:19:53.200 --> 00:19:59.279
Well you you are going pretty fast, and the asphalt is pretty hard, I have to say.
00:19:59.279 --> 00:20:00.799
So it's no joke.
00:20:00.799 --> 00:20:03.599
Not very pads either.
00:20:03.759 --> 00:20:04.640
Not very pads.
00:20:04.640 --> 00:20:07.839
I have a my wife is much more indulgent though.
00:20:07.839 --> 00:20:09.920
She never said and she never said stop.
00:20:09.920 --> 00:20:12.319
So I just I just kept going at it.
00:20:12.559 --> 00:20:17.680
You're just a you're a tough person, then you know, but I don't know, but I was tough.
00:20:17.759 --> 00:20:18.559
I'm more fearful.
00:20:18.559 --> 00:20:22.319
Like if I don't get back in my bike, I would be one of those people, right?
00:20:22.319 --> 00:20:24.160
So that's good motivation.
00:20:24.160 --> 00:20:26.240
All right, all right, gotta get back on subject.
00:20:26.240 --> 00:20:40.160
All right, so Bobby, if you could do this, if you could actually take your sheet of paper, I reminded of a colleague of mine who was in I was at engineering.com, I was the design guy, he was a manufacturing guy.
00:20:40.160 --> 00:20:47.519
And every day practically he would tell me, hey repenter, I can take this drawing into the shop and they'll manufact.
00:20:47.519 --> 00:20:51.519
I'm sorry, take a sketch into the shop and they'll manufacture it for me.
00:20:51.519 --> 00:20:56.559
I I just put a quick sketch, give it a couple of dimensions, and there you go, right?
00:20:56.559 --> 00:21:00.640
So I want you to next time we talk, or maybe now you can even do that.
00:21:00.640 --> 00:21:01.599
I can I do that now?
00:21:01.599 --> 00:21:02.880
Can I show hey Jim?
00:21:02.880 --> 00:21:04.400
Not his real name.
00:21:04.400 --> 00:21:08.000
Maybe hey Jim, somebody can actually do this now.
00:21:08.000 --> 00:21:09.680
They can actually give a sketch.
00:21:10.000 --> 00:21:10.319
Yeah.
00:21:10.319 --> 00:21:11.920
We we how about I do this?
00:21:11.920 --> 00:21:15.599
How about I show you a pre-recorded video and then we can go into a live demo as well.
00:21:15.599 --> 00:21:17.039
Everything will break.
00:21:17.039 --> 00:21:19.359
Something that works and some that breaks.
00:21:19.359 --> 00:21:21.039
So um yeah, let's try that.
00:21:21.039 --> 00:21:25.839
Um, but I love your idea here of just uh just that one example alone.
00:21:25.839 --> 00:21:32.960
We've heard that's just one of many examples where many people have a 2D drawing and they just need to get 3D.
00:21:32.960 --> 00:21:40.400
They will actually bring these to machine shops or send it through to Zometry, Proto Labs, yeah, wherever you want.
00:21:40.400 --> 00:21:42.960
And they'll say, Do you have a 3D file?
00:21:42.960 --> 00:21:47.359
And the answer is, of course, like no, obviously I don't, because otherwise I would have given that to you.
00:21:47.359 --> 00:21:54.400
But hey, I have here I have a fax, or I have a photo, or I have a scan, or I just drew this by hand and I just need a promade.
00:21:54.400 --> 00:21:58.640
It'll say, Okay, well, give us a few days and we can make a CAD model and then we'll quote it for you.
00:21:58.640 --> 00:21:59.599
But many people won't even.
00:21:59.599 --> 00:22:10.480
Quote apart because they know that just the effort alone for them to spend on CAD, and you know, machine shops are not stocked up with engineers that are waiting for a job to come through so they can do CAD.
00:22:10.480 --> 00:22:12.480
No, they actually run production parts.
00:22:12.480 --> 00:22:16.960
They only want to run production if they're not turning spindles, they're just not making money.
00:22:16.960 --> 00:22:20.720
So they don't want to bother with CAD, but they could at a cost.
00:22:20.720 --> 00:22:24.240
So four to five days later, it'll kind of say, Hey, we we made a model.
00:22:24.240 --> 00:22:26.319
Do you want to purchase this part?
00:22:26.319 --> 00:22:27.519
They'll say, sure.
00:22:27.519 --> 00:22:37.279
And now they're gonna pay you know another hundred, three hundred, four hundred dollars for their time that someone had to spend not running G not running Fusion or not programming a machine, but now they have to do CAD.
00:22:37.759 --> 00:22:37.920
Right.
00:22:38.319 --> 00:22:39.200
This happens quite a bit.
00:22:39.200 --> 00:22:43.839
We know about 20% of inbound to machine shops falls under this category of fabricated parts.
00:22:43.839 --> 00:22:47.839
And many times fabricated fabrication quotes come in with no 3D.
00:22:47.839 --> 00:22:57.200
So here's something that we have this on maybe on LinkedIn or maybe on Instagram, but uh it's just something that we ran at the automate show in Chicago.
00:22:57.200 --> 00:22:58.720
This was about two months ago now.
00:22:58.720 --> 00:23:07.839
Automate show is all about you know manufacturing automation, but it was a great place for us to launch our product, and we had people make AI, make parts using AI in our booth.
00:23:07.839 --> 00:23:11.279
Tormach is one of our partners, and we had a machine from them.
00:23:11.279 --> 00:23:22.160
People came, draw, drew something, put it through inference, we put into Fusion, G code, post to the Tormach, make the part, and the end of the day they have something that they actually made with AI.
00:23:22.160 --> 00:23:27.200
So I'll just play this all the way through and um drag your part into here, generate.
00:23:27.200 --> 00:23:30.880
We have some rather cheesy SimCity sort of messages there.
00:23:30.880 --> 00:23:32.160
The part is made.
00:23:32.160 --> 00:23:34.319
You could inspect it, download as a STEP.
00:23:34.319 --> 00:23:38.319
Sorry, this is all quite quick, but download as a STEP, open this infusion.
00:23:38.319 --> 00:23:40.559
You know, we we're not automating G code.
00:23:40.559 --> 00:23:43.440
We could, but there's plenty of places that do that already.
00:23:43.440 --> 00:23:44.559
The part is there.
00:23:44.559 --> 00:23:52.799
We already have our template set up, so we just basically throw it in there, generate G code, simulate, make sure we aren't gonna crash anything like crazy.
00:23:52.799 --> 00:23:56.319
Uh post the G code to our Tormach.
00:23:56.319 --> 00:23:58.720
And again, Tormach's been a great partner for us.
00:23:58.720 --> 00:24:06.240
There more CAD in the world means more parts to be made, which means more revenue for a CNC maker like Tormach or really anybody else.
00:24:06.240 --> 00:24:08.319
But you know, we threw the part in here.
00:24:08.319 --> 00:24:14.960
Um uh we're doing all this on the fly so that we could try to get people apart at the end of the day.
00:24:14.960 --> 00:24:18.559
And um, we just make the part right there on the spot.
00:24:18.559 --> 00:24:29.680
And through this, people were like, wow, I could actually have a part I designed with a sketch, and I could have this in as quick as you could click a mouse to get the G code over to the machine.
00:24:29.680 --> 00:24:34.000
So that's just like a really quick like this is why we're making AI CAD.
00:24:34.000 --> 00:24:46.799
It's not so that people could sit, um it's not it's not so that people could spend more time in CATIA or have to figure out how to, you know, run a machine.
00:24:46.799 --> 00:24:52.480
We just want people to get CAD because with the STEP file, if they have that in their hands, they could go to a number of places to get it made.
00:24:52.480 --> 00:24:57.759
They could work on it themselves, they could send it to on-demand manufacturing, they could 3D print it.
00:24:57.759 --> 00:24:59.599
So the STEP file is really key.
00:24:59.599 --> 00:25:02.319
Uh, you know, I don't think a lot of people care about the workflow.
00:25:02.319 --> 00:25:12.400
And no most just like you said, you know, it's like a it is as much of a is almost more of an art than a skill to build a CAD model really well.
00:25:12.400 --> 00:25:24.960
Um feature recognition is great for step files, but uh sorry, close that window, but it is a whole art to create a feature tree that is actually updatable if you were to change one feature.
00:25:24.960 --> 00:25:31.599
Automotive parts are a good example of sheet metal stampings, and this is more than just you know press break parts.
00:25:31.599 --> 00:25:38.079
These are like there's a lot of curvature in a frame rail, and there's a lot of curvature in a door panel or a inner panel.
00:25:38.079 --> 00:25:53.440
Um to really build those pieces, and I've built many of them in my career using KTIA, is 100% a black art of maintaining your feature tree so that you can update anything and it doesn't crash your model.
00:25:53.440 --> 00:25:56.559
And you know, I'm sure you understand that as well as a lot of the viewers do.
00:25:56.559 --> 00:25:58.240
So we don't want people to care about any of that.
00:25:58.240 --> 00:26:00.720
We just make the step file directly itself.
00:26:01.039 --> 00:26:02.079
Yeah, that's good.
00:26:02.079 --> 00:26:02.720
That's good.
00:26:02.720 --> 00:26:09.200
So do you feel like there's are you competing with anybody else to make specifically what you're doing?
00:26:09.200 --> 00:26:11.039
Or are you are you it's a great question?
00:26:11.119 --> 00:26:14.000
We're competing for sure in the industry of AI CAD.
00:26:14.000 --> 00:26:18.880
I had an investor ask me once if I felt that the industry was tainted.
00:26:18.880 --> 00:26:20.960
And I said so.
00:26:20.960 --> 00:26:35.519
I think even they understood that you know, AI CAD companies um that are working in this domain of text-to-CAD or um image to CAD, they face struggles with people maybe not necessarily being able to trust the output.
00:26:35.519 --> 00:26:41.440
Yeah, I think it will take time for the AI CAD startups of the world to gain industry trust.
00:26:41.440 --> 00:26:54.240
But I think it's largely because a lot of AI CAD companies are claiming to do way more than they're capable of presently, and also claiming to do way more than they ever will be able to in the future using LLMs alone.
00:26:54.240 --> 00:27:06.480
And for I think for me it's a bit hurtful because you know we're trying to approach this from a very different perspective of going straight from the embeddings of a neural net to a STEP file.
00:27:06.480 --> 00:27:26.240
This bridge between AI LLMs and CAD as agents, or LMs as writing CAD query code or open cascade code, there's so much that gets lost in that translation because um, you know, one CAD agents are not always able to understand how to build a good model.
00:27:26.240 --> 00:27:31.839
They could do okay, but when you get to like a sheet metal, like an automotive stamping, like deals off.
00:27:31.839 --> 00:27:44.799
If you look at CAD query as a way to write code to make STEP files, and some of our competitors do that, and they do decently at that, you're fundamentally quite limited in the terms type of geometry that you can create as well.
00:27:44.799 --> 00:27:50.960
So you'll never be able to get an automotive stamping out of a lot of these uh kernel, primitive kernels that exist.
00:27:50.960 --> 00:28:05.519
So by working on the STEP file directly, because the STEP file is the graph that is the actual output, we know we at least have the availability of as infinite types of geometry that we would not be restricted by a kernel.
00:28:05.519 --> 00:28:09.039
And we're also not restricted by someone having access to CAD software.
00:28:09.039 --> 00:28:15.759
So our population is much bigger of users, as well as the geometry that we can create is much wider as well.
00:28:16.160 --> 00:28:16.480
Yeah.
00:28:16.480 --> 00:28:24.240
I'm glad to see somebody with actual CAD experience could mention Katiya, but sounds like you also know several others too.
00:28:24.640 --> 00:28:27.039
I've dabbled in others, kind of like you said before.
00:28:27.039 --> 00:28:30.000
I have not spent enough time to really get very good at things.
00:28:30.000 --> 00:28:36.400
And they um NX is most similar to Katia, I think, in the capabilities because of the kernel.
00:28:36.960 --> 00:28:40.640
If you learn Katiya, you probably learned the uh one that's the hardest one.
00:28:40.640 --> 00:28:43.839
So you could probably have an easy time learning anything else.
00:28:44.160 --> 00:29:01.599
Well, it's it's easy, but it's actually very frustrating to use other CAD systems because Katiya is so powerful, just like NX is very powerful, as you know, to really design and handle the very complex curvature, it's not just curvature, complex calculations required to make fillets.
00:29:01.599 --> 00:29:07.039
I heard on someone else's podcast the other day, and and uh this is coming from from Carl Bass.
00:29:07.039 --> 00:29:13.119
He said something, uh, the question was like, is there any real difference between you know CAD programs these days?
00:29:13.119 --> 00:29:14.240
Or are they all the same?
00:29:14.240 --> 00:29:17.759
And I think Carl said something like, Well, fillets are pretty much just fillets.
00:29:17.759 --> 00:29:22.240
So he didn't think that there was any real significant difference between CAD programs.
00:29:22.240 --> 00:29:31.119
And yes, a fillet is a fillet, but actually the way that the engineer has to operate the software can be very different from CAD system to CAD system, as you know.
00:29:31.119 --> 00:29:33.519
And so that's why Katia, I think, is just king.
00:29:33.519 --> 00:29:34.720
For me, it's king.
00:29:34.960 --> 00:29:39.920
I'm surprised you didn't make a like a gold plated and hang it around your neck like a medallion.
00:29:40.960 --> 00:29:47.519
We we did if you uh poke around in other places on LinkedIn, you'll see um we we actually did that.
00:29:47.519 --> 00:29:49.359
Uh we had several gold chains.
00:29:49.359 --> 00:29:50.640
I have some here.
00:29:50.960 --> 00:29:57.599
I may have to mail you one, and people were making oh so but you want the but you want the gangster put that on.
00:29:57.920 --> 00:29:58.160
Correct.
00:29:58.160 --> 00:30:02.160
If I can find a photo of one and I have some, I'll I'll email them to you.
00:30:02.160 --> 00:30:14.640
But no, we absolutely did that because we wanted people to do really interact with what AI CAD is and how it might do something physical in their life and not just exist as a CAD, CAD alone.
00:30:14.640 --> 00:30:18.640
Um that really is the whole point of CAD, anyways, to get a real part in your hand.
00:30:18.640 --> 00:30:20.960
And so we want people to have that opportunity.
00:30:20.960 --> 00:30:26.559
Now I know you you asked if we could test this a little bit, and uh I think we probably should.
00:30:26.559 --> 00:30:30.160
You know, this is available online, you know, we're not hiding anything here.
00:30:30.160 --> 00:30:32.079
Uh it's just an MVP.
00:30:32.079 --> 00:30:34.799
You can sign in with your email address.
00:30:34.799 --> 00:30:37.920
Now you're welcome to try this.
00:30:37.920 --> 00:30:40.400
We have credits for free.
00:30:40.400 --> 00:30:43.759
Uh, you get four generations for free as it stands currently.
00:30:43.759 --> 00:30:50.640
And this is the like the administration view, so you know some buttons you wouldn't have access to, but let's go to like this demo one.
00:30:50.640 --> 00:31:00.160
Um one of the ones that is interesting, and again, this is like where we bridge the gap between artistic intent and an engineered part.
00:31:00.160 --> 00:31:01.920
We'll get better at this over time.
00:31:01.920 --> 00:31:07.680
But I took this photo at a steakhouse restaurant and because my nephew said, Can you make that?
00:31:07.680 --> 00:31:09.200
I'm like, let's try it right now.
00:31:09.200 --> 00:31:10.240
We tried it on my phone.
00:31:10.240 --> 00:31:17.759
So, you know, this is one of the examples of how AI CAD doesn't have to be in SOLIDWORKS or whatever your CAD is.
00:31:17.759 --> 00:31:25.519
It doesn't have to be run on a computer that could compile, or not compile, but run Python scripts that make a STEP file for you.
00:31:25.519 --> 00:31:27.920
You know, you know, we run this on cloud.
00:31:27.920 --> 00:31:30.559
You could do it on your phone, iPad, computer.
00:31:30.559 --> 00:31:31.759
You just need to be able to take a photo.
00:31:31.759 --> 00:31:36.079
And this will take a few minutes, so we could just continue chatting as a yeah.
00:31:36.319 --> 00:31:45.119
Now are you gonna are you gonna cut it up into those, you know, those stakes pieces, you know, like how they end up with dotted lines to be necessary?
00:31:45.680 --> 00:31:46.720
We definitely could.
00:31:46.720 --> 00:31:54.000
We didn't, I don't know if we could do that yet, but that's an interesting very interesting case of uh good industry to get into, I suppose.
00:31:54.160 --> 00:31:54.880
Meet cutting.
00:31:54.880 --> 00:31:55.519
All right.
00:31:56.000 --> 00:32:05.599
But we found yeah, we found that we had some requests from uh future prospective customers of like, can you take the water jet shop, water jet laser cutting shop?
00:32:05.599 --> 00:32:09.359
They said we get drawings that are just hand drawn all the time.
00:32:09.359 --> 00:32:11.680
And some of them are pieces of art.
00:32:11.680 --> 00:32:29.759
And um an artist will come to us and say, you know, we want this cut out of a slab of of iron for like this ironwork, or you know, you see like you know, this uh quartz steel kind of fancy looking stuff, and they'll just come with a drawing of that they did by hand and say, Hey, how much will this cost?
00:32:29.759 --> 00:32:33.519
You know, the programmer they'll look at it and say, like, that's like a couple hours.
00:32:33.519 --> 00:32:35.519
So it'll charge them for that time.
00:32:35.519 --> 00:32:40.799
And sometimes the the job is a no-quote because it's like that's just we don't have time for that.
00:32:40.799 --> 00:32:44.079
Sometimes it might cost too much and someone walks.
00:32:44.079 --> 00:32:56.960
If someone could just actually dump that into a CAD, dump that in as a 3D file, then they could program it on their OMAX right away or put it straight into their plasma right away and then go cut it.
00:32:56.960 --> 00:33:21.839
So what would be easiest for that situation is if they could just put the 2D drawing into something that's a better version of Transfigure, get a STEP file out, and if it's got a few inaccuracies, at least with a STEP file, they could just you know round a corner, cut something out, add this or subtract that, and get it, they get like 90% of the work done, and then just go straight to cut it on their laser or water jet.
00:33:21.839 --> 00:33:27.119
And this is not a perfect cow, but actually did a decent job because again, we have to really bridge the differences.
00:33:28.240 --> 00:33:32.160
It was able to uh satisfy uh separate it from its background and everything.
00:33:32.160 --> 00:33:35.279
So this is yeah, this is really pretty good, I would say.
00:33:35.279 --> 00:33:43.359
Of course, not anatomically perfect, and certainly not uh 3D, but uh but still that's I'm impressed.
00:33:43.359 --> 00:33:44.160
That's pretty good.
00:33:44.160 --> 00:33:51.039
And yeah, so I was able because of the contrast, it was had a it had a somewhat, I wouldn't say easy time.
00:33:51.039 --> 00:33:57.519
I don't want to downplay what you what you did here, but because it's good contrast, I was able to make that outline out, right?
00:33:57.839 --> 00:34:01.599
Yeah, so you know we we are not perfect, so it is AI.
00:34:01.599 --> 00:34:04.480
So as the disclaimer goes, check the output.
00:34:04.480 --> 00:34:06.880
And if it doesn't work the first time, try it again.
00:34:06.880 --> 00:34:24.880
You know, we we've benchmarked this against other things like just running an agent using Chat GPT or Claude or Gemini, or actually Gemini is actually maybe one of the worst, but between Chat GPT or or Claude, this stuff is possible as well.
00:34:24.880 --> 00:34:36.880
But the amount of tokens spent just to think through the problem exhaust sometimes your daily or at least your hourly allotment, and sometimes you're never sure like will I actually get a part at the end or will I run out of tokens here?
00:34:36.880 --> 00:34:44.559
You know, each generation right now we charge a customer like 50 between 20 to 50 cents, depending on what credit package that they bought.
00:34:44.559 --> 00:34:48.960
But you always know, okay, uh, we'll get an output because we're not working with LLMs.
00:34:48.960 --> 00:34:53.599
So our pipeline doesn't spend a lot of tokens thinking, it's really just doing.
00:34:53.599 --> 00:34:55.760
I think that's the best way to describe it.
00:34:55.760 --> 00:35:00.400
Uh, but yeah, that's this is one just a fun example of something that's that's a good one.
00:35:00.639 --> 00:35:01.440
Can I ask you something?
00:35:01.440 --> 00:35:02.960
You have a thousand credits there.
00:35:02.960 --> 00:35:05.119
I think my the free account comes with four.
00:35:05.119 --> 00:35:06.639
So what's four credits worth?
00:35:06.639 --> 00:35:08.239
These said 50 cents.
00:35:08.480 --> 00:35:08.719
Sure.
00:35:08.719 --> 00:35:13.199
So you know, if we go up here to the add credits button, I apologize, this is a little bit.
00:35:13.199 --> 00:35:16.960
Oh, I could blow it up here, but we have a small current reload.
00:35:16.960 --> 00:35:21.039
Um, think of this as the like the least value at McDonald's.
00:35:21.039 --> 00:35:25.440
$10 gives you 20 generations, so you get 20 Big Macs for 10 bucks.
00:35:25.440 --> 00:35:32.239
But if you do the supersize me, um, you know, we're overly generous and we'll do for 200 bucks, a thousand generations.
00:35:32.239 --> 00:35:36.079
So this is 20 cents per generation here, how we've set it up now.
00:35:36.079 --> 00:35:38.480
And this is 50 cents per generation down here.
00:35:38.480 --> 00:35:44.880
Um, I will let you and you know the future viewers know that we are about to roll out an upgrade to the pipeline.
00:35:44.880 --> 00:35:48.159
Uh, we're gonna test pricing a bit with a slightly upgraded model.
00:35:48.159 --> 00:35:52.639
Um, okay, way more capable, but it will definitely burn more of these credits.
00:35:52.639 --> 00:35:57.440
So the pricing will per generation will depend on the model in the very near future.
00:35:57.599 --> 00:36:03.119
Um you're trying to make you're trying to make money now, you're subsidizing this still.
00:36:03.119 --> 00:36:06.880
That's not the true cost of the AI, though, is it?
00:36:07.199 --> 00:36:10.559
Well, you know, including everything for sure.
00:36:10.559 --> 00:36:19.119
You know, we're a startup, we're burning cash like crazy, but we we are committed to to never selling credits, selling generations themselves for a loss.
00:36:19.119 --> 00:36:23.840
I think that's very important because it's it is pretty hard to raise prices.
00:36:23.840 --> 00:36:30.480
It's it's because it is so difficult to raise prices, um, we've made sure to never sell this for a for a loss.
00:36:30.480 --> 00:36:36.719
I think there's some other companies out there that are probably selling things for a loss somehow, but that's we never want to do that.
00:36:36.719 --> 00:36:40.400
We'd rather have people wait it out and then be prepared to pay.
00:36:40.400 --> 00:36:40.880
Yeah.
00:36:40.880 --> 00:36:43.920
So we actually make some money per generation here.
00:36:43.920 --> 00:36:48.079
Uh not very much, but we actually do make money because we've calculated the cost very carefully.
00:36:48.079 --> 00:36:51.280
We uh we know exactly how much compute we use per generation.
00:36:51.280 --> 00:36:54.719
We're not arbitrarily just churning on tokens until we run out.
00:36:54.960 --> 00:36:59.599
So it's this this cow, that XF piece, that's what when I create that, that's one generation.
00:36:59.840 --> 00:37:00.960
That's one generation, okay.
00:37:00.960 --> 00:37:02.400
Uh one credit right now.
00:37:02.400 --> 00:37:03.519
Yep, that's right.
00:37:03.840 --> 00:37:09.440
I'm glad to see this because no, no, no time have I really been made aware of my token use.
00:37:09.440 --> 00:37:15.840
And all the time I question, I uh I say in trees, like how many trees have I just burnt down, you know?
00:37:15.840 --> 00:37:16.480
Sure.
00:37:17.360 --> 00:37:21.360
So yeah, here's another one that I could just run idly in the background.
00:37:21.360 --> 00:37:26.239
It might fail, it might succeed, but we could just kind of continue chatting as as we're going on here.
00:37:26.239 --> 00:37:31.920
Um, you're right that a lot of our competition, and there's definitely some out there, which is good.
00:37:31.920 --> 00:37:33.679
We're not the only crazy ones doing this.
00:37:33.679 --> 00:37:39.840
We're maybe a little crazier than most, but our competition oftentimes has uh you know some kind of daily limit.
00:37:39.840 --> 00:37:54.320
Some other people are doing maximum generations, but I've used other people's platforms and actually found it pretty frustrating to almost get to the point of maybe there's a part that's about to come out, but then it actually just aborts and says, okay, you ran out of your tokens.
00:37:54.320 --> 00:38:07.599
And I have to say, unfortunately, I think because a lot of AI CAD out there right now is so below my personal standard, and even this is like a little bit for sure, it's like below my personal standard of how I run Katea, clearly.
00:38:07.599 --> 00:38:15.199
I I have I find it very difficult to pay for subscription to a service that actually is quite bad.
00:38:15.199 --> 00:38:26.719
Um I'm not saying that we're somehow so much better than everybody else, but we didn't want people to feel like the only way to try this out is to subscribe, and then you got to worry about your credit card being charged month to month.
00:38:26.719 --> 00:38:31.280
That's why we said just let people purchase like a small bag of tokens and they could do what they want with it.
00:38:31.280 --> 00:38:36.639
They don't expire, we'll upgrade over time, and they'll get even more value out of the tokens as time goes on.
00:38:36.639 --> 00:38:42.079
Uh, but the whole subscription model for AI CAD is is total hot garbage right now.
00:38:42.079 --> 00:38:43.440
So it doesn't make any sense.
00:38:43.760 --> 00:38:44.079
I like it.
00:38:44.079 --> 00:38:48.239
I like it because you get to try it, you get what you pay for, and you're not locked in.
00:38:48.239 --> 00:38:52.480
And you know, I don't I don't want to pay for shit that doesn't work anyway, right?
00:38:52.639 --> 00:38:55.760
So yeah, but ten dollars we found is a good thing.
00:38:56.000 --> 00:38:58.000
I'm not calling your stuff that at all, by the way.
00:38:58.000 --> 00:39:04.000
I'm just saying the CAD, the text to CAD stuff that I worked, which is gives me preposterous results, right?
00:39:04.000 --> 00:39:07.840
That's like I would never subscribe, but I would try it once and then never use it.
00:39:07.840 --> 00:39:10.400
I definitely wouldn't subscribe to it, right?
00:39:10.639 --> 00:39:13.440
Yeah, so that that's that's the uphill battle that we all face.
00:39:13.440 --> 00:39:19.440
Um that's why one VC asked me if the world is if the AI CAD world is tainted, and it absolutely is.
00:39:19.440 --> 00:39:29.280
Yeah, but we've never operated in this world of LLMs as replacing someone's ability to do CAD in a software program or write CAD as code.
00:39:29.280 --> 00:39:37.920
By working on this in the within the neural net itself, one, it's the only way to actually oops, that one failed.
00:39:37.920 --> 00:39:39.199
I'll try to run it again.
00:39:39.199 --> 00:39:41.920
That's the only way for us to actually scale this up.
00:39:41.920 --> 00:39:49.039
There's only about one, I found there's only about one million CAD licenses that are purchased or annually renewed every year.
00:39:49.039 --> 00:39:58.000
That's a pretty small population, but it's also pretty interesting because like there are only one million copies of CAD out there that make everything in the world around us.
00:39:58.000 --> 00:40:10.400
This whole my whole backdrop and you know, the chair you're sitting at, or the tables that we work on, the clothes that we wear, they're produced or made of machines that existed as a CAD design somewhere in the world.
00:40:10.400 --> 00:40:17.039
So a very small number of people actually make everything in the world that we exist in, but that population is still pretty fixed.
00:40:17.039 --> 00:40:21.920
There are many other people in the world that would love to just draw apart and receive something in the end.
00:40:21.920 --> 00:40:34.480
So we service those customers, and um, I find that to just be like a way bigger surface area problem population that we could reach of people that want to build and repair the world, more than just engineers like to make stuff.
00:40:34.480 --> 00:40:36.000
So that's kind of how we see it.
00:40:36.079 --> 00:40:40.320
Um tell me this about about the uh abilities so far.
00:40:40.320 --> 00:40:44.079
That's right now it works at, we see that it works at sheet metal.
00:40:44.079 --> 00:40:48.239
I imagine the plans are to make it work on other other stuff.
00:40:48.239 --> 00:40:50.480
Also, what if what about can it do?
00:40:50.480 --> 00:41:02.719
Do you want to to eventually handle sheet metal as design, uh sheet metal parts as designed, or are you going to stay in the flat sheet metal has to be unfolded and flat?
00:41:03.039 --> 00:41:10.159
Oh, so so this one if it works and doesn't crash out, this one is a is a bent piece of sheet metal, kind of like a Z bracket was.
00:41:10.159 --> 00:41:35.679
You know, there one of the things that we do, and I don't think this is this is definitely not a trade seeker or anything, but you know, we capture attributes of the part as JSONs, like as structured data, so that later if we want to modify Ben radius, as your example you brought up, and quantize the output so that we know, okay, like 0.6 or 3, or even just making the inside radius the same as the thickness, we could we could absolutely do all of that stuff now.
00:41:35.679 --> 00:41:38.800
And those things are oh, I'm a little disappointed here.
00:41:38.800 --> 00:41:44.559
Actually, I think that there may be an issue in our pipeline, but let me try this again one more time.
00:41:44.559 --> 00:41:52.880
But you're you're absolutely right that there are certain rules of manufacturing that we even we don't want to break in our model.
00:41:52.880 --> 00:42:02.559
So if there's no bend radius information shown, it might be the best option just to make all the inside bend radii one thickness, and that's probably fine for most people.
00:42:02.559 --> 00:42:05.920
And if it's not, you can put it into SOLIDWORKS.
00:42:05.920 --> 00:42:12.400
You could do the sheet metalization, and you could just modify all of them to whatever your tooling size is in your shop as well.
00:42:12.400 --> 00:42:23.039
So um, so yeah, there's a there's a lot of things that we need to improve in this pipeline as it stands now, but it is just an MVP and it's not our foundation model that we're training now.
00:42:23.039 --> 00:42:30.320
Our foundation model is actually something altogether different that is much more difficult to copy than what we've done here.
00:42:30.320 --> 00:42:32.800
So that's uh that is coming soon.
00:42:33.039 --> 00:42:34.639
What do you think the problem might be here?
00:42:34.639 --> 00:42:36.960
Is it said it's something in the pipeline?
00:42:36.960 --> 00:42:39.360
So we use it for your cloud services.
00:42:39.599 --> 00:42:39.920
Yeah, yeah.
00:42:39.920 --> 00:42:49.519
So I'll just say that right now we are in the process of upgrading with this update, and our developers might be doing something in the back end that is preventing this from working.
00:42:49.519 --> 00:42:50.079
Okay.
00:42:50.079 --> 00:42:52.800
But um, you said you made that Z bracket before.
00:42:52.800 --> 00:43:00.079
Um, the Z bracket is an example of a part that is a has a bent shape and it could be unfolded in you know, Xometry.
00:43:00.079 --> 00:43:02.400
You go to Xometry, it'll unfold it for you.
00:43:02.400 --> 00:43:06.320
There's some other uh sheet metal shops like uh Oshcut.
00:43:06.320 --> 00:43:08.559
OSH Cut will unfold it.
00:43:08.559 --> 00:43:12.719
You could send that to SendCutSend and they'll make it for you.
00:43:12.719 --> 00:43:18.639
Although SendCutSend seems to prefer 2D drawings instead of 3D files.
00:43:18.639 --> 00:43:27.760
SendCutSend is an interesting example of a company that stands to benefit from having more 3D and less 2D.
00:43:27.760 --> 00:43:35.280
In fact, on their website, I noticed recently you can send in a 2D drawing of a sheet metal bracket and it'll make a 3D file for you.
00:43:35.280 --> 00:43:36.639
Oh, what's coming here?
00:43:36.639 --> 00:43:38.800
Okay, so we made a thing.
00:43:38.800 --> 00:43:39.920
It is not perfect.
00:43:39.920 --> 00:43:46.320
I mean, even you know, my sketch it has this kind of like lean leaning tower of pizza to it.
00:43:46.320 --> 00:43:48.400
So um what you get with a hand.
00:43:48.400 --> 00:43:49.119
With units.
00:43:49.119 --> 00:43:49.840
Yeah, exactly.
00:43:49.840 --> 00:43:56.079
So we do size it to if there's information on the drawing, we will do our best to size it proportionally.
00:43:56.079 --> 00:43:59.519
Um and we have this banana for scale.
00:43:59.519 --> 00:44:05.280
So we can understand, okay, is this like part a millimeter big, or is it maybe close to the size of a banana?
00:44:05.280 --> 00:44:08.239
But that's just clever.
00:44:08.239 --> 00:44:10.000
But yeah, we do all this with math.
00:44:10.000 --> 00:44:11.920
This is not a LM approach.
00:44:11.920 --> 00:44:21.840
We're not burning tokens on Claude or Gemini or ChatGPT because we know that in the end, an LLM operating in CAD is possible.
00:44:21.840 --> 00:44:28.239
And we do we could do some of that, but that's definitely not the foundation model approach that we're taking because it's just not skip.
00:44:28.239 --> 00:44:30.000
So yeah, this anyway, some simple example.
00:44:30.000 --> 00:44:30.880
You know, interesting.
00:44:30.960 --> 00:44:32.719
So it's all it's all equations.
00:44:32.719 --> 00:44:35.119
This is not uh correct.
00:44:35.119 --> 00:44:37.519
This is not a neural network behind this.
00:44:37.920 --> 00:44:41.440
It is we do this is half AI and half code.
00:44:41.440 --> 00:44:50.960
So what we do here is we'll actually take the 3D representation and then we algorithmically create this CAD model as a STEP file directly.
00:44:50.960 --> 00:44:53.039
And we again we don't use an LM to do that.
00:44:53.039 --> 00:45:08.639
We're just using our engineering brains and we wrote code that is a little bit brittle, but it is minimally functional because we don't want to pay Claude or pay Anthropic or ChatGPT for the benefits of using LM here because we know that's not the end game.
00:45:08.639 --> 00:45:18.320
Um but this part where we take the 3D representation and write the STEP file, that is the part that AI has not solved yet, but that is what we're working on in our foundation model.
00:45:18.320 --> 00:45:28.400
And once we've done that, and we've finally unlocked Terminator from Skynet, and um that's where I believe we should be a little bit worried about AGI.
00:45:28.400 --> 00:45:43.599
Uh but yeah, you know, I just want to kind of run through maybe just the rest of this here because we if we have a little bit of time, um and you know, to show, you know, with that bracket that we just downloaded, I guess it's probably this is my account.
00:45:43.599 --> 00:45:47.119
Xometry is a great place to order parts, very convenient.
00:45:47.119 --> 00:45:48.159
The price is right.
00:45:48.159 --> 00:45:50.400
Oh, I I guess I did one here before.
00:45:50.400 --> 00:45:54.079
Um, let's try to let's just try to grab this one.
00:45:54.079 --> 00:45:56.639
Bracket.step is the one I just downloaded.
00:45:56.880 --> 00:46:00.960
No, it did not say $11,860, did it?
00:46:00.960 --> 00:46:01.599
Did you make it?
00:46:01.760 --> 00:46:03.280
That's for one platinum.
00:46:03.280 --> 00:46:05.599
Sorry, that's sorry, that's for like a thousand.
00:46:05.599 --> 00:46:06.000
Sorry.
00:46:06.000 --> 00:46:06.800
Oh, okay.
00:46:06.800 --> 00:46:07.039
Okay.
00:46:07.039 --> 00:46:08.960
You'll see when this comes up what it says.
00:46:08.960 --> 00:46:11.119
So, you know, if you want to buy one, okay.
00:46:11.119 --> 00:46:16.320
I'd charge 165 bucks for doing a pain in the butt one part too.
00:46:16.320 --> 00:46:23.599
You know, if we change this to, I don't know, a thousand or something, you know, obviously we're we're paying a lot more, but now it's like nine, ten bucks a pop.
00:46:23.599 --> 00:46:24.320
Oh, okay.
00:46:24.320 --> 00:46:27.199
And you know, I could go and order this part right now.
00:46:27.199 --> 00:46:36.159
And so that's that's kind of showing like the end-to-end goal here of why AI CAD is because you can get a part very quickly in your hands.
00:46:36.159 --> 00:46:41.519
I could order this to arrive by September 9th and pay a ten thousand dollar expedite fee.
00:46:41.519 --> 00:46:45.119
But you know, these things are all possible once you have that STEP file.
00:46:45.119 --> 00:46:45.440
Right.
00:46:45.440 --> 00:46:50.320
If you don't have a STEP file, uh, because you know Xometry is just gonna sit on it and or they would no quote it.
00:46:50.320 --> 00:46:50.880
Yeah.
00:46:51.280 --> 00:46:51.440
Yeah.
00:46:51.760 --> 00:46:58.960
So I'm guessing that Xometry does not have this app this uh capability.
00:46:58.960 --> 00:46:59.840
Uh, are you?
00:46:59.840 --> 00:47:05.199
Um I'm guessing also uh putting it maybe putting a thought in your head.
00:47:05.199 --> 00:47:06.559
Oh yeah, I had the thought already.
00:47:06.559 --> 00:47:08.320
Are you thinking Xometry is gonna buy you?
00:47:08.320 --> 00:47:10.239
Is that the uh exit strategy here?
00:47:10.480 --> 00:47:16.639
I would love to entertain their offer, but ultimately it'll just depend on timing, right?
00:47:16.639 --> 00:47:20.159
I think it depends on on where we're at as a company.
00:47:20.159 --> 00:47:23.760
Um you know, we're we're raising a seed round right now.
00:47:23.760 --> 00:47:30.239
If Xometry were to buy our company and give us the backing to push us all the way through, sure, I would definitely consider it.
00:47:30.239 --> 00:47:35.519
I don't think that they will, but also I don't necessarily want to have someone dig their claws in so quickly.
00:47:35.519 --> 00:47:39.199
You know, we've also had people like Autodesk ask us about what's up.
00:47:39.199 --> 00:47:43.760
Autodesk would eat us and swallow us whole and spit us about, I think.
00:47:43.760 --> 00:47:50.000
Um but on the end, Autodesk is a great company and they make great product and they're huge, you know, a huge, well-respected company.
00:47:50.000 --> 00:47:59.840
But I think um we need to make sure that in order to get this job done and not have someone kill the company, I think we need to do this ourselves for the time being.
00:47:59.840 --> 00:48:02.880
We want to be like the chat GPT of CAD.
00:48:02.880 --> 00:48:08.320
I'm sure many other startups say the same thing, but in order to do that, we have to build our own foundation model.
00:48:08.559 --> 00:48:12.480
So yeah, yeah, definitely just gonna take a foundation model.
00:48:12.480 --> 00:48:16.800
And by the way, a word of advice, don't trust Autodesk.
00:48:16.800 --> 00:48:21.519
Because they have their own, they're making their own foundation model as as we speak.
00:48:21.519 --> 00:48:30.079
And uh and I don't I think the only reason they might buy anybody that else is making a foundation model is is to get them out of the running.
00:48:30.400 --> 00:48:34.159
Uh that I see the world very similarly, and I wish them luck.
00:48:34.159 --> 00:48:39.440
But like I mentioned, you know, Autodesk, sure, their market cap is huge and they're well deployed.
00:48:39.440 --> 00:48:45.840
But you know, I would say Autodesk is also not very much adopted at the automotive OEMs.
00:48:45.840 --> 00:48:54.239
And the main reason is like you cannot make hoods and bumpers and headlights and all these things that you know class A surfacing really demands.
00:48:54.239 --> 00:49:00.800
So Autodesk can't even make automotive, really automotive parts, but also that population is really fixed.
00:49:00.800 --> 00:49:09.679
We don't want people to have to buy an Autodesk Fusion, even like $50 subscription to Fusion just to use an agent.
00:49:09.679 --> 00:49:20.159
We aim to give people of any background with a little bit of pocket change, an ability to make a part that they could use to improve their own life, regardless of pad.
00:49:20.159 --> 00:49:20.400
Yeah.
00:49:21.679 --> 00:49:23.199
Is it just you doing this right now?
00:49:23.440 --> 00:49:24.320
It's just me.
00:49:24.320 --> 00:49:25.360
Okay.
00:49:25.360 --> 00:49:28.719
We have a lot of contractors helping us, so I'm very grateful for that.
00:49:28.719 --> 00:49:34.800
But um, no, I am uh I'm a solo founder and uh uh it's uh that's what I do.
00:49:35.039 --> 00:49:37.440
No, no, it's it's good, very good so far.
00:49:37.440 --> 00:49:40.880
I expect greater things, and good luck on your seed round.
00:49:40.880 --> 00:49:44.800
I think it's a very good idea, and uh you convinced me anyway.
00:49:44.800 --> 00:49:52.800
Of course, I'm not an investor, but I think it needs to be done by people who are not just fantasizing about CAT.
00:49:52.800 --> 00:49:57.119
They don't know what CAT is or what it does.
00:49:57.119 --> 00:50:01.280
And you know, as an engineer, I trust 3D geometry.
00:50:01.280 --> 00:50:06.880
I don't trust any kind of shape that I've seen that AI has made for me so far.
00:50:07.199 --> 00:50:09.360
Yeah, we're all at the same starting point though.
00:50:09.360 --> 00:50:14.239
You know, I won't I don't want to throw so much shade on all of our competitors because we're all trying to figure it out.
00:50:14.239 --> 00:50:28.880
But I absolutely think that um, yeah, because I've just worked in the industry of automotive OEM and a little aircraft OEM for for so long, you know, the pain is very real to me of why CAD, good CAD really matters.
00:50:28.880 --> 00:50:40.719
Um and the ability to talk shop with the guys on the factory floor to really understand their needs as an ultimate customer, because they're the ones that have to deal with the CAD that you know we make.
00:50:40.719 --> 00:50:45.199
The ultimate customer is the guy on the shop floor um that has to make the part.
00:50:45.199 --> 00:50:53.199
Of course, the end user as well, but you know, the direct customer is the fabrication engineer, and um we know how to talk shop with them.
00:50:53.199 --> 00:50:55.679
So I'm wearing my my shop shirt for you today.
00:50:55.920 --> 00:50:56.960
That's right, that's right.
00:50:56.960 --> 00:50:58.000
We saw that, yeah.
00:50:58.000 --> 00:50:58.800
Yeah, very good.
00:50:58.800 --> 00:51:00.079
Good promotion.
00:51:00.079 --> 00:51:05.119
So uh that was good work on getting the uh partnership with the machining company.
00:51:05.119 --> 00:51:06.159
Tormach.
00:51:06.480 --> 00:51:07.280
Tormach, yeah.
00:51:07.280 --> 00:51:12.400
Yeah, more cad means more parts, which means more NC or fabrication equipment.
00:51:12.400 --> 00:51:13.360
Yep, absolutely.
00:51:13.920 --> 00:51:23.199
All right, Bobby, thanks for thanks for reaching out in the first place on LinkedIn, and then secondly for being a guest who do our part to get the word out there, yeah.
00:51:23.360 --> 00:51:24.559
I appreciate your time.
00:51:24.559 --> 00:51:32.079
It was great talking with you, and um I have I love the people that you bring on for your interviews.
00:51:32.079 --> 00:51:44.320
We it's just such an awesome range of like the really OG CAD people that were in it before V4, even, as well as us new folks that are trying to do this whole AI CAD thing.
00:51:44.320 --> 00:51:47.280
So great uh great job covering the full spectrum.
00:51:47.599 --> 00:51:48.800
I I love doing it.
00:51:48.800 --> 00:51:49.840
All right, very good.
00:51:49.840 --> 00:51:50.719
All right, Bobby.
00:51:50.719 --> 00:51:51.039
Thanks.
00:51:51.039 --> 00:51:52.159
Have a good rest of the day.
00:51:52.159 --> 00:51:52.800
Thank you.
00:51:52.800 --> 00:52:00.400
Okay, thank you for listening to FoDES, the Future of Design and Engineering Software show, brought to you by ENGtechnicaa.
00:52:00.960 --> 00:52:05.599
I hope you have learned of a new application or technology that will help you with your job.
00:52:05.599 --> 00:52:16.880
If you have an application you think would be of interest to other engineers, please let me know by emailing me at roopinder at enchtechnica dot com or message me on LinkedIn.