Welcome to NVIDIA's AI podcast. Picture this.
It's 3 a.m. You find yourself in a tattoo parlor But none of the dragons, flaming skulls or gothic font lifestyle mottos you see on the wall seem like something you'd want on your body.
You know what I'm talking about. We've all been there.
So what do you do? You turn to AI, of course.
Our guests, Goran Vukšić and Dennis Jensen, are part of the development team behind Copenhagen and New York City-based TattooDo.com. which uses deep learning to help anyone find the tattoo of their dreams.
Goran and Dennis, welcome. Thank you. Thank you very much.
So tell us about this. Like, Describe it sort of at a high level and let's dig into kind of what's actually happening under the hood because I went to your site and What you do is you type in some keywords.
In my case, it was ocean and whale and homeward.
And it comes up with images of tattoos that... correspond to those words and images that you might like.
But what's the point and how did you guys build this thing?
Goran, walk us through it. Tattoo is the number one place in the world to get your tattoo inspiration.
Now we have almost 1 million users and we have over 350,000 tattoo images. uploaded by users and it's a pretty huge database where you can find a lot of inspiration and style.
And beside the web, which you mentioned, we also have the mobile apps.
Once you register in the mobile app, you can use our Discover tool to search through different styles and find the artists and shops and different tools and such. but a really interesting part of our application is our home feed because we We track what the user is doing in the application.
And if you are liking some dragons and following artists that are that are good in making those dragon tattoos then we will try to offer you through the home feed more of those motives you look for, or more of some styles?
It just occurs to me, here we are talking about about AI and tattoos, and if you had asked me if those two worlds would come together, my answer would be no.
We might be talking about AI and Mountain Dew or something that's kind of prototypically computer science-y.
So for you guys, and I mean this personally, which came first, the tattoos or the AI?
And do you find yourself... cultural anomalies?
Are AI people getting cooler or are tattoo people getting more into computer science?
So, I had the opportunity to join this company almost two years ago. where we have amazing co-founders and an amazing team behind Tutudo. that were also aware about that, okay, this is the first time to do pop culture getting and as it's more and more accepted people like it when it's when it's online and it's new and because we're we're We are kind of the pioneer of digitalizing this industry, and it's a huge opportunity.
It's a huge uncovered market. it wasn't like immediate apparent that AI would fit this in.
But Gwen and I had a fun idea for this hack day we have every quarter at the company where developers can kind of do exactly what they want for a whole day and then show it to the other guys and then it would eventually possibly be evaluated to go into the roadmap. into the actual product later down the line.
So it actually came a bit out of the blue.
However, Goran has a lot more experience in both AI and also augmented reality.
So he kind of got the initial idea and asked me to join in on it.
Well, but let me ask you this. Yes, it's true that tattoos have become the sort of broad-based cultural phenomenon.
Everybody's got tattoos and they have them all over the place.
But there's also that – so there's that, but then there's also the kind of cultural phenomenon of – computer science.
And it strikes me that you don't find developers covered in tattoos all the time, or maybe you do.
And maybe that's what's changing. And maybe that's why these two worlds.
When I found it out, I was really interested. to get this job because both me and my wife are fans of tattoos.
We have one friend that is a tattoo artist.
And like that's always been somewhere around, you know, something interesting and I think I got my first two like some five years ago uh-huh yeah maybe Maybe it was a bit late, but I said, okay, when I'm doing it, I'm doing something serious, so I uncovered whole my back.
With what? It's demon wings with barbed wire around it.
So when I got the chance to jump on board with Tudu, I was really happy and I'm still really happy about it because It's something interesting.
It's some kind of art. Well, and so how do you and Dennis, I have to ask, I assume that you have tattoos as well.
Yes, I got my first tattoo on my back when I was 18.
And then I actually only got the remaining two. tattoos while working at Tatuto.
So the first year I worked there, six months after I started, we released the first real version of the mobile application.
And we had a huge party where all the investors came in.
Amit James came in. He said two different kind of people.
And, uh, We actually set up a tattoo studio at the office and we were able, everyone was able to kind of get us a two of the one.
So I got my second tattoo there, and then the following year, The first one I got was a traditional tattoo on the left side of my stomach. which is a wolf howling in the sunset.
The year after, I got one by the same person on the other side of my stomach, which is a panther.
Well, so I – and I can see this. The mobile app gets exactly to my hypothetical example in the introduction.
You can carry these images with you everywhere.
But what is going on in a sort of deep learning sense?
And You say that it's users who upload the images.
Is that where you're training the data, essentially?
So how we initially gathered the training data for all this is by user tagging when they upload. their photos, they have an opportunity to add hashtags and mentions and and the location on the body for this particular tattoo.
And so what we did was to aggregate a huge list of these common hashtags that we like to categorize and then we use those hashtags to map to these common things that we'd like this AI to learn. and then we basically went through all our pictures and tried to categorize them to prepare the AI for this particular training.
We boiled it down to 11 styles, which we then queued through the whole digit system on an Amazon instance.
How tricky was it to teach the system to actually learn and be accurate about when I type in a key word and whether that's a a place in the body or a thing to get it right.
So we basically had our community manager do a lot of research to find these different classifications and boil down to two types of classifications, which was Motives and Styles.
And Styles was the most interesting one to start out with.
So he did a lot of initial manual work to to boil down a lot of these styles and also because styles are very much subjective.
So he did a lot of initial work, But other than that, we could programmatically go through all our pictures and sort them and prepare it for the training.
I would like to add from those 350,000 pictures we have, Only 100,000 we were able to take and associate with specific styles.
And that's why we needed some system to classify the remaining pictures.
And when we are able to do that, We can offer our users like better content.
We can offer them better suggestions. So we were thinking what to do here.
Let's try to teach the neural network to recognize those styles.
Our base was these 100,000 images that we used to train the neural network.
I don't know if that part of like if you use tattoo do to kind of like narrow down your choices, but how do you see this working? for users like how do you how do you see them you know their behavior or their use cases How does it work for them?
And then do you see a decision point where they go from looking at images to like, okay, Now we're going to book an appointment with a tattoo artist that I found through tattoo do as well.
Yeah. So, so a quick thing, um, other than this is like a whole social media with all this inspiration of tattoos the next major point This whole application is to be able to provide a direct link to a shop or artist where a member can book an appointment and already now you can actually get in contact with a shop or an artist as a member through the application or or through the web platform.
So we're also working towards making it much easier for a member to get there to two of the dreams with the artist that they really like by providing them a direct channel to actually communicate with the artist I mean, Goran, is that kind of chain of events and AI is kind of this engine that helps connect those people? those products and services.
Are we going to see more of these kinds of systems in the in our lives, you know, whether it's tattoos or otherwise.
And do you think that it's got You know, a future that we haven't even seen yet.
Well, definitely. We'll see it. And, like...
Now when we trained our neural network, what we did with this model, we used it with new iOS 11 that was just about to come out and use it with core machine learning.
And like you probably know and follow those mobile trends that AI is more and more coming to Syria and all other countries. stuff.
And yeah, I believe it will give us more and more relevant suggestions so it's not just some you know dummy search on a website when you get and get some results but it's actually tracking of user behavior and trying to figure out what user really wants, what he needs, and offer him relevant content based on that.
And I think, interestingly, it also gives users kind of a glimpse into what they don't want.
So, for example, when I did my search of whales and...
And, you know, the word homeward, you start to think like, oh, you know, maybe I don't want to wail on my body or, you know, maybe I mean, for me, it sort of felt like.
I thought that was original, but clearly, you know, there's thousands of people who have the word homework on their, you know, tattooed on their bodies.
I think it works both ways. Yeah, that's true.
Like, if you are searching for some... some really common motif like skull or dragon or whale or something like that.
But if you're looking for some specific style on the other hand, you can find some artists that are really good at that style and decide what kind of style you want and then talk with that artist to make you some custom tattoo.
Pardon the interruption. But we have a favor to ask.
Leave us a review on iTunes, Google Play Music or whatever your favorite podcast platform of choices.
It helps more people find us. As always, thanks for listening.
Now back to the good stuff. Can it be predictive?
Do you guys have a sense of kind of where tattoo culture is headed next based on...
I don't know, the interaction and kind of what your system is learning?
Obviously, for every submission we get to this, to the platform, we have the opportunity to learn something from this in an AI sense.
Keep feeding... keep feeding this model with new data and actually just keep improving the success rate of the whole AI.
For now, it's actually quite open what this can actually be used for like really, really, you know, other than just what it's used for now.
But I definitely think that the opportunity is going to open way more in a few years. really really you know all new and there's not a lot of people that knows about the internal works of it and it's really really amazing and really impressive but But there's actually no limits where this actually can be used for because it's business intelligence and it's It's a way of figuring out what your members and your customers want before they actually know it themselves.
Right. And so it's this combination of images and, you know, keywords and images and ideas in some sense and people, right?
This combination. and you could apply those elements to other kinds of business cases or, you know, artistic cases or services, et cetera, however you want to describe it.
Exactly. Well, I, Let me ask you guys both as developers, I mean, how did you get into the world of AI and machine learning? in the computer science world and you had an interest in the AI side of things.
Were you in the art world and you had an interest in like the computer science side of things?
I mean, how did it work for both of you?
Goran, let's start with you. well like I'm really interested in augmented reality for some years now and like Usually in my free time, I always try to do some cool things, some weekend projects for myself or to entertain my son. when Apple decided to go out with iOS 11 it was basically two main new things.
It was core machine learning and ARKit. I really wanted to try this core machine learning, so I spoke with Dennis.
We have a huge database, let's try to make something make some sense out of it and like we had few discussions and then we we gave it a go All before was just like, yeah, okay.
I understand how it works, but never had the chance to do something this big.
Well, and Dennis, what about for you? I'm actually completely green in for this area, disregarding the last three months.
So because I've always been just a backend developer, really interested in web development and software architecture, et cetera. and I moved into a more DevOps field, setting up our infrastructure, etc.
So Goran kind of just came to me saying, okay, we have this idea.
If I was able to spin up the machines that we need, we need the software installed on it, obviously, a GPU powered instance.
I asked if I was up of doing it and I was, Actually, other than we're going to show me the projects he has done previously, this is the first time I actually had hands-on experience on this whole project. the whole flow and how this whole works.
But to be honest, At this point, still, the more I read about it, the more stupid I feel because there's so... many new concepts.
Every time you learn about a new concept, there's four types of flavors of that concept.
But that's just a typical learning curve.
It's complex things. For those people who are thinking of doing a project like this or an app and like, hey, we want to get our feet wet or get our entire bodies wet.
Like what's your advice and what did you learn along the way in these last three months of of getting this from, you know, idea to to a reality?
Well, uh, I think the most important thing is to like sort out your data for learning because like, I think, I believe if data is right if it's sorted out correctly that and if you have like nice amount of images you can train from like you can probably do it for different things.
Dennis, anything that you learned that you would give people as a piece of advice?
Yeah, for sure. Like the number one thing you kind of need to have this going is a data set that you could start out training with.
But the most important thing is just to like, figure out what is it you want to you know you want to do just on a really low level basis do you want to detect something in an image or do you want to you want to predict something in another sense, but there's a lot of guides out there I spent a lot of time on Cora learning about the different frameworks for learning. going through the cafe that we used initially, that's Cafe 2, the TensorFlow.
All of them starts out with, find a data set you wanna try to play around with, and then just sever through Google, look at different guides, try to, Try to just get it up and running, not necessarily understanding all the concepts because then you'd never get going, but just try different guides and just Yeah, just try to get something up and run, something to work and then kind of move backwards and try to understand more specifically what's actually happening.
What's the next tattoo? Do you have one in mind?
Are you using tattoo do to narrow it down?
I like... black and gray styles. It'll probably be something like that.
And like I i will probably do the sleeve on the left hand next but like yeah i i still have to figure out what exactly and of course the todo is giving like uh nice suggestions so right dennis what about you I am pretty confident that I'm going to get something on my sternum by the same guys that did my other tattoos. which I indirectly found through Titudo, but not through the app, actually, though.
I'm just going to start getting my own stomach filled up.
That's actually the first plan. Right. Wow.
So you know the artist, not necessarily the image, and you'll go from there.
Exactly. Goran Vucic and Dennis Jensen, Tattooedoo.
I will think about my own tattoo, and I'll get back to you guys when I decide on which whale I'm going to get and wear.
Thank you. Thank you, and thank you for having us.
