GenAI Meets Celebrity: Inside Cameo’s Journey from Startup to Stardom

This week, David is joined by Dom Scandinaro, CTO at Cameo,  Cameo, the platform that allows users to send personalized videos from celebrities and creators. Scandinaro gives an in-depth overview of Cameo’s rapid growth from a simple idea into a robust and deeply personalized fan experience platform with the digital infrastructure to match. 

Join us as we discuss:

• How Cameo’s tech stack has evolved as the company expands globally

• Why leaders need to have an in-depth understanding of their team’s abilities and limitations

• How Cameo is harnessing Generative AI for new products and platform features.

David Joy:

What is up, everyone? And thanks for tuning. In today's episode of the Big Ideas in App Architecture Podcast, we speak to Dom Scandinaro, who is the CTO of Cameo, best known for its platform where users can request personalized video messages from a wide array of celebrities. Dom and I talk about Cameo's tech stack and how it has been architected on the cloud for scale.

We learn about the engineering culture at Cameo and get into some exciting new GenAI features his team is working on. Pump up the volume, and enjoy listening in on this episode.

All right, Dom, welcome to the Big Ideas in App Architecture Podcast. How are you doing today?

Dom Scandinaro:

I'm doing really well. How are you?

David Joy:

I'm good. You're joining us from Chicago, from the headquarters of Cameo. Is that correct?

Dom Scandinaro:

Yes, that's right.

David Joy:

It's a really good-looking room there. I can see it's a warehouse kind of space. Is that the case?

Dom Scandinaro:

Yeah, it's definitely very, very modern. Kind of warehouse-y vibe. Exposed ceilings. The exposed vents. Brick walls. That sort of thing.

David Joy:

Very cool. I mean, I'm really excited to have you on today. I know we had this very interesting way of connecting with each other when I saw you presenting it, an AWS part to Summit, and I was ... While you were talking, sent you a LinkedIn message saying, "I absolutely love what you are talking about. Would love to dig more into that. Would you love to come on the podcast?" And you're so generous. Thank you for saying yes to come on.

Dom Scandinaro:

Of course. Thanks for having me. I'm happy to be here.

David Joy:

Brilliant. One of the reasons why I was really excited to talk to you was obviously Cameo is a company that I've heard about, and I've seen a bunch of ads and some really cool things that happen on social media around the company. I was really excited to see somebody from that space talk about technology. And what really intrigued me was how Cameo as a company is getting shaped up and is building this amazing platform for customer experience or user experience that provide them a touch of celebrity in many ways.

And when I was researching the company, I kind of went to see who the founders were, stocked a little bit about the story, and got to know that you guys started off in 2016, where one of the founders essentially tried to get a video of an NFL player to a friend, and that sparked the idea, and Cameo was formed, right? And so, it's brilliant. I wanted to get into that a little bit and get to know you.

Before I butcher your introduction, why don't you let the people know about who you are, a little bit about the company, and your role at Cameo?

Dom Scandinaro:

Yes. My name is Dom Scandinaro. I'm the CTO at Cameo. I've been in that role for I think about six months now, but I've been head of engineering for about two and a half years. And I've been at Cameo for a little over five years.

Been in engineering leadership for the last 10 or 11 years of my career at this point, but actually got the opportunity to join Cameo as a individual contributor. Senior software engineer. Got in really early post-series B as the team was kind of split scaling. And there wasn't a management position opportunity at the time. Got a chance to kind of join as a individual contributor and just work heads down, mostly, for five or six months, which was a really refreshing kind of experience to just tackle some problems and take a beat to not have to worry about other people and their performance and motivation and leading a team and all that sort of stuff. But before I know it, those opportunities came full circle for me at Cameo as well.

And Cameo ... You kind of nailed the start of the business. But at the heart of it, we're a two-sided marketplace. We connect fans to a marketplace of creators and celebrities. We have about 50,000 celebrities and creators on the marketplace today. Our core product is a personalized video shoutout message, which hopefully most of the audience is familiar with but the kind of default use cases ... For my wife, for her birthday coming up later this year, I might get her a video from maybe a star of a reality TV show that we're watching together at the time or something like that to just wish her happy birthday, touch on a few things that are really specific to her and her life, her business, and some of the progress and milestones that she's been able to overcome in the last year, and they create a video and send it back to you. We're really just kind of the glue between those two sides of the marketplace.

David Joy:

Well, that's really awesome. I mean, I've seen a bunch of videos around. When I was researching, I saw that you started ... You were at around 40,000, but it seems you've grown very rapidly over the years and now at 50,000. That's really a great scale of growth and millions of personalized videos as well.

Tell me a little bit about when you joined or when you started. Where was Cameo technology wise? And how has it changed from a tech stack point of view as the business grew?

Dom Scandinaro:

I'd say that it's both changed a ton and very little at the same time. We had a technical co-founder, Devon, from day one. He still here in the business. He runs our design team today, and he wrote all the original code. We started out with React on the front end, Note on the back end, and MongoDB as our database.

I think we're storing all the videos and images and stuff in S3 from the start within AWS as well. And we didn't have an app when we started. We have this belief in fighting for simplicity and not solving problems we don't have. And I think the videos originally were ... The orders were sent to the celebrity side of the marketplace via Telegram or Facebook Messenger, and then they would record a video and send it back to the bot, and that would get it uploaded.

We didn't have an app in the early day.s and when we decided to build one out, we started that and React native just to keep the same TypeScript React stack in place. And that has scaled up to be able to fulfill the business that we have today.

The ways that we have evolved are more around the hosting and scalability of things. We started on Heroku as just a really simple note app. When we outgrew that, we moved to Containers, and today we run on Kubernetes and AWS. But at the heart of it, the tech stack is still really, really similar to what it was in the early days.

David Joy:

Very good. And so, when you joined the company, were you initially taking care of the app? What area were you initially focused on?

Dom Scandinaro:

My very first thing was that the whole business is kind of built on push notifications. Somebody comes to the site. They book a video request. And something that's unique to Cameo is that every single celebrity creator on the platform has our app on their phone. But if push notifications aren't working, they're not going to know that you booked a request from them. Some people are really busy, and they're probably opening the app every day anyway, but most people are getting booked a couple times here, a couple of times there. They get really ... Maybe they have an article posted about them. They get 20 bookings in one day, and then they go two weeks without any.

But our push notifications were not working very well. They were not reliable. My first task was kind of digging into what was going on there, evaluating some different providers that we could potentially switch to, and ultimately just hardening the push notification infrastructure end to end. That scaled from the app to also include the back end. But the first few teams I worked on and the first couple of teams I led were all very app specific at the time. A lot of work on the celebrity side of the app and the interface they have for recording and uploading the videos and also a lot of work just on the early consumer version of the app that we had in the app store.

David Joy:

That's awesome. I'm thinking when you were saying the push notification is so critical for your business. I mean, celebrities being who they are, they have all these things going on, a busy schedule, and then they're adding something like this to their part of the story of what they want to do. Getting them the ability to be aware of, "Hey, you got to do this," is very important.

I'm going to ask a very curious question just out of suddenly [inaudible 00:08:35]. Who is the most popular celebrity on Cameo right now?

Dom Scandinaro:

It is very incredibly variable. Definitely changes kind of week to week, month to month. But Brian Baumgartner, Kevin Malone from the Office, is definitely the OG most popular person on Cameo of all time. He's still incredibly popular right now.

But it's August. It's actually fantasy football season at Cameo. We're a very seasonal business. John Gruden. Antonio Brown. Scott Hanson from the RedZone. The three of them are just kind of getting booked hundreds of times, literally, a day to announce the draft order of fantasy football drafts that are coming up. I actually just did mine for a league I'm in last night. This is the heart of the season. We had a Scott Hanson video to announce our draft order that was really, really incredible, who was actually in the Red Zone studio when he recorded it, which was super cool to see. Kind of did a backstage tour of things. Those folks are the most popular right now because it's August, and NFL kickoff's here in a week and a half.

But right from there, we'll go into holiday season. And we'll have creators dressing up in a Santa Claus outfit and getting booked for $10, $20, $30 that ... We'll have hundreds of Santa Clauses on the platform getting booked and doing videos. There'll be people dressed as elves and other holiday-related characters. We'll go right from fantasy football season into the holiday season here.

David Joy:

Right. I mean, that was good that you brought up because I didn't think about that perspective. It seems like it's very much on where the fans are at that point of time that defines who they want to reach out to and what kind of messages they want. It's definitely very seasonal, it seems.

When I was researching, I found that as you're scaling, and scale has become so critical for you, you went from just being a North America company to now looking as a global company and trying to grow globally.

How has that affected your tech stack and your seasonal behavior that affects your traffic?

Dom Scandinaro:

I mean, really from the early days of having to migrate off of Heroku and into AWS, a lot of the motivation for that was scale. Our business has been inherently spiky really from the beginning because so much of it is based on virality around somebody joining Cameo. If a big kind of A-list celebrity joins the platform, there'll be 10, 20 articles published the same day about them. A surge of traffic will come with each one of those articles getting posted and shared out.

Similarly, somebody will record a really awesome globally interesting video. There was John Gruden did one, I think, for a fantasy football draft, maybe roasting somebody that lost the year before or something like that, and it went hyper viral a couple months ago. And that was just millions and millions of hits directly to his profile page because this video was getting posted out on Twitter and then reshared and reshared kind of all over the internet.

We've had those moments since the early days. Our infrastructure has been very elastic since early on because we don't want to pay for there to be 50 or hundreds of pods running all the time just in case one of those surges happens. But instead, we want to be elastic and be able to scale up and down as needed throughout the day on everything from the website to the app to the API back into the database. We've had to face a lot of those challenges since early.

One of the ones that was most interesting to me, personally, is we had ... When we had a little bit larger of a team, and we were working on a lot of product lines at the same time, we had two newer products. One was called Fan Clubs. It was kind of a PG version of OnlyFans, and it still exists today as a product called Follow. And the other was called Cameo Calls. That one has since been sunset, but it was really popular during stay-at-home orders during the pandemic. And that was being on a FaceTime call on your iPhone with a celebrity. And both of those products were very timely. And people used them together.

Tom Felton was one of the people that was most popular on Cameo calls, and he would post out to all of his followers through his Fan Club on Cameo that he was about to go live on a Cameo call. He would go on Instagram live and tell all of his followers on Instagram he was going to go live on Cameo calls, and then we'd get the surge of traffic all within 10, 20, 30 seconds, not spread out over several minutes, of people just trying to be one of the 10, 20, 30 people that got a ticket and got in line to be part of his ... Have that one of those one-on-one calls with him.

And that brought a whole new level of scale to us because previously, when something would go viral, one person sees it on the news, and another person sees the article get posted on a Twitter. People don't necessarily have notifications turned on for those things that they're going to hit instantly. With this product, they were literally subscribed to say, "I want a notification when Tom Felton goes live on Cameo calls." And when he went live, they would hit that notification instantly. And this spike of thousands of people surging to the app at the same time was something that we were prepared for scale, but that was a little bit more direct and really, really specific in the same matter of seconds than what we had seen before where it'd maybe be spread out for minutes or hours.

David Joy:

Well, that's brilliant. That scale is ... You're talking about a hockey-stick curve that just happens immediately, right? I believe that you transitioned, from what you were saying, from Heroku to Amazon EKS. And so, how do you deploy today? Is multiple regions? Is it a regional cluster of EKS pods?

Dom Scandinaro:

We're all in one region. We have fall tolerance and disaster recovery backups and plans to be able to scale over to a second region if we need to. But, no, our scaling is all within multiple zones in the same region.

David Joy:

Same region. That's amazing. Well, I mean, those are the kind of things that I love to talk about. That's brilliant.

I wanted to come back to the scale part in a bit. I know, when you and I were talking, and we cannot escape this at all right now, this is, I would say, the pool where we get submerged is generative AI. And you are really excited about some of the stuff that you're working on. Some of the things you even brought up during your AWS presentation and the panel that you were on.

Tell me a little bit about what you guys are doing with generative AI. And what are you personally excited about?

Dom Scandinaro:

Yes, GenAI, the topic of 2024.

We have four really exciting things that are noteworthy and worth covering, kind of going from the past to the present to the future. The first is Cameo Kids. About a year and a half ago, we launched a new product called Cameo Kids. Essentially, if you think about the 50,000 celebrities and creators we talked about, those are obviously all humans. They have the app on their phone. They have real things coming up in their life day to day.

And Cameo Kids is the opposite. They're animated characters. The voices of those characters are GenAI-powered voice models. Those are trained by the actual voice actors that power the characters on the TV shows or movies or what have you that they're known for. And they come into a studio. They read hundreds of lines, and a voice model is created from that. At the end, we have a text-to-speech model. We can put in a sentence of a personalized video request, and we'll get back out that sentence in the voice of the character based on that model.

We have some different nuances for 3D characters versus 2D characters and how the animations get created, but ultimately that all gets stitched together with FFmpeg, and we're able to create dynamic videos that we can present to the customer.

There is a human in the loop moderation there. The IP holders themselves are very involved in the process. They're watching the videos. They're helping tweak and phonetically pronounce some of the more difficult names to make sure the video is really, really special and personalized and authentic to the way that the character would say it themselves, but they approve the videos, and they get sent right back to the customers, all just generated by technology. That's really cool.

It's a little bit limited today just by the comfort level of the IP holders and what they do and don't want their characters to say and what comfort they have around what they want them doing. Love to see where that can go in the future as the technology gets more and more acceptable. But that's a big thing that we did previously.

Jumping to this year, we launched review summaries. If you're familiar with Amazon.com's product detail page, they've got hundreds and thousands of reviews on many of their products, and it's really hard to kind of sit through all of them. The star reading is all most people look at. They launched the review summaries where AI is really great at summarization. It's really easy to have it summarize text. It's something it's really, really good at. Doesn't hallucinate much in that area.

Just sending all of your product reviews, in our case, talent video reviews, into the AI model and saying, "Summarize these," putting some prompt engineering to teach it Cameo's brand voice and what we do want it to focus on what we don't want to focus on works really well out of the box.

And what we found is unique to Cameo is that people are buying these videos from people they love more than anything in the world. Their favorite sports star. Their favorite actor or actress. And sometimes they'll get back a video that's just not that great, but they'll still give it five stars. Maybe in the text, they'll actually mention, "They actually mispronounced my wife's name," or, "The video was only 15 seconds. I wish it was much longer." That stuff is something we're able to really surface through the review summary. We can give the customer a much better expectation of exactly what they're going to get.

Some people do record shorter videos. Some people record really long detailed videos. Some people want that. Some people want the one that's just short and to the point. But just giving people all that information upfront is a really cool feature.

And what we didn't expect is that the celebrity side of the marketplace actually loves them. We see them screenshot them on their profile page, share them out to social, and be like, "Look at what the AI is saying about me," because it really pulls out the stuff that they do super well and complements them in a really, really nice way, which they're proud of, which is really cool to see.

David Joy:

We get right back to today's episode in just a moment. But before we do, I want to let you in on a secret. Big ideas, like those you hear on this podcast every week, don't need big databases to start with. With CockroachDB Cloud, you can bring your apps to life quickly and without upfront cost architected on the same resilient distributed SQL platform that industry leaders trust with their mission-critical workloads. Sign up for free and start bringing your big ideas to life today at cockroachlabs.com/bigideas.

I mean, this is amazing. This whole large corpus problem has been there for a long time, and I think what AI has allowed us to do is sift through this very quickly and summarize it.

I let you complete the other two, but I had a few questions on these. I can come back.

Dom Scandinaro:

Sure. The third one, excuse me, is really similar to the second. We launched a new version of our celebrity profile page couple months ago in an AP test. And a big component of that that's primarily highlighted at the top of the page is the description of the celebrity themselves. We've always had a bio that you could write anything you wanted in as a celebrity or creator on the platform. And some people write a really great description of what they're famous for off of Cameo, what they want to be doing on Cameo, but way more people just write something like, "Hey, I really wish you booked me on the web instead of the app because I lose 30% to Apple in the app," or, "I only want to do business requests at my business price," which is relevant to 5% of people visiting Cameo that are actually here for our B2B product.

What we're able to do with this is we're able to summarize what they're famous for off Cameo and what they like to do on Cameo. And we tried a bunch of different methods. AI is really good. It knows the history of what these people did on the internet, but you quickly realize most models haven't been trained or retrained in the last six months. A lot of people on Cameo are famous for something that happened yesterday, not famous for something that happened six months or a year ago.

And also, the information changes, right? You're on one sports team. You get traded to another. You're in one TV show. You're now working on a second. The data moves quickly. And what we found is that nobody knows the celebrities and creators on our platform better than our customers. And they're actually already writing all of these details into the requests when they book the videos on our platform. They're writing that they love them because of their time on the New England Patriots, or they really liked their performance in this specific TV show playing this character name.

We're able to take the public requests that are booked on our platform and put them through a model and, again, with some prompt engineering and brand voice and that sort of stuff, have the AI describe what they're famous for off-platform and ultimately what kind of videos they do like doing on Cameo. Because some people are great at birthday requests. Some people do awesome roasts, but not everyone is known for all of the occasions we have. Giving that information to the customer as well is really, really great.

Now, we have both. We have the AI version at the top of the profile and then the chance for you to still highlight some specific details you want to add yourself just below that. We have the best of both worlds, which is really cool. But that's also just a summary use case at the end of the day.

And then jumping to now, we're working on ... And this is part of what I was talking about at the AWS event where we met. We're working on a chat-based discovery project. We have 50,000 celebrities and creators. We currently let anyone in the world join. David, you're welcome to after the show to join as Cameo talent and set your price, and anyone will be able to book you. But with that comes a tremendous problem of continuing to make our browse and search functionality better and better and better and continue to collect more and more structured data on the people that are on Cameo so that we can get people that are here to purchase a video to the right person easily.

And today we're not great at that. One thing we're trying to do to overcome that is to create a chat experience where, if we're able to detect that maybe you're having trouble finding the right category of talent with the right specific talent for the person in your life, toss you into a conversational interface, ask you in a back-and-forth conversation, "Are you buying for yourself or someone else? What's your budget? What occasion are you booking for?" And then tease out some of those details that they might not be thinking of themselves if they're buying for somebody that they're a little bit less personally familiar with.

What's sports team do you love? What city does the person live in? What TV shows or movies do they like? And then based on the categories it's trying to match you to and all of those little tidbits we pulled out of it, we can get you to the best category and then filter the talent within that category based on the things that you told us in the chat.

We were building a POC of that earlier this year in partnership with a company called Loco that I was on stage with at that AWS event. And the POC was compelling enough that we've moved into a build-for-production phase where right now we're working on the final steps of some iteration and hope to have this live to customers this fall on Cameo's website.

David Joy:

I mean, this is amazing. All of these projects that you mentioned. I mean, these are the kind of things that I feel people are able to leverage AI in a very accurate way. I loved all the four projects that you mentioned.

The last one, I particularly ... Going back, I love because I, as a user, would be like, "Oh, I want to give my wife a celebrity to come and talk about," and I have a celebrity in my mind because that's what I'm thinking. But then when I start putting some information about her, I'm pretty sure it's going to recommend somebody else in that list which are like, "Oh, maybe this one's better," or something like that. And we don't think like that naturally. And you intrinsically thinking about that customer experience is something that's coming out in this conversation where I feel you're thinking about what might be more valuable to the user who's coming to Cameo. He might have a single thought process, but maybe we should expose him to some other experiences that he can or some other celebrities. That's really cool.

Dom Scandinaro:

Totally. Very, very related to that, when you do have a celebrity in mind, we don't have 10 million celebrities. We have 50,000. People are often, "I want to get a video from Taylor Swift or Lizzo or Tiger Woods," and it's easy for us to index maybe those top 10 or 20 people and say, "If you say Tiger Woods, then golf. If you say Taylor Swift, then pop music."

But one thing that AI embeddings really I hope will do really well ... They look like they're going to excel in our testing so far is the ability to say, "Okay, you said Tiger Woods. Tiger Woods is a golfer. The closest thing in an AI embedding model to golfer is the Cameo golf category." There are subcategories in there that are very popular of people that are on the PGA Tour. Recommend the PGA Tour. And that's something that just being able to do at scale, especially with the small team that we have at cameo today in 2024, is just something that doesn't feel possible without this generation of technology. That's just really, really exciting as well.

David Joy:

I think these are really cool use cases, even the summary use case, and pretty much it seems ... Just to summarize the couple of use cases that you mentioned, it's essentially a building a rack system where you're using an existing train model, which might not be completely accurate, but then you are using some of the data that your users are putting and then combining that with the AI model.

If I can ask, and if you're allowed to say it, are you more on the open AI side or on the cloud side or on the Mistral side? Or you're building your own models?

Dom Scandinaro:

For each of these projects, we've kind of done our due diligence and tried different models on different cloud providers. What we've found is that we've been happiest with the Claude family of models, and we specifically have been running and building these projects on top of AWS Bedrock. We just feel really comfortable with the security and legal models and frameworks that they have in place. The shared security model, I think, is the right nomenclature for the AWS policy that's on top of all their services that, if you've been in the industry like we have for the last 15 years, we've been building on AWS for that entire time, and they've had the same model in place.

This is an oversimplified statement. But if you're comfortable storing your data in a database that's managed by AWS, then you should be comfortable putting your data through a model. Obviously, with the exception of PI and that sort of stuff, putting your data through a model that's hosted on AWS Bedrock ... It's the exact same agreements and security practices that are in place there. They abstract the underlying model completely away. None of your data is available to any of the vendors providing the models for training or anything like that. It's a really streamlined way for us to be able to meet our legal and security requirements internally and still be able to build on this next-gen technology quickly.

But we do try out different models from different providers. And for some projects, one model seems to work better than others. We're always kind of flexible along the way.

David Joy:

That's what I've noticed, too, and especially this year has been, I think, a ... Nobody talks about it much, but it's sort of a Claude year because they kind of surprised everyone with the Sonnet and the Haiku model. And I went and tested them. And I was extremely surprised on how consistently good they were. And one of my other issues that I was having with some other models, without naming them, obviously not Claude, were that they would give me responses, but they're not consistent all the time and had high levels of hallucination and then had to tweak some or add some extra information.

Dom Scandinaro:

I think Claude prides themselves on accuracy, lack of hallucinations, being appropriate, avoiding topics that are controversial, so it feels a little bit easier. I think, with any model, if you put the right guardrails and stuff in place on your own side and within your code, you can make the stuff work, but I think what they give you out of the box has a lot of the bells and whistles already, and then you don't have to go add yours on top, even though within AWS Bedrock they obviously give you the ability to set those guardrails up and protect yourself anyway.

David Joy:

I've been enjoying reusing Bedrock. I have a couple of bugs that created against the team, myself, that we are still working through. But generally speaking, I mean, it's fascinating how quickly the cloud providers are able to bring all these capabilities into the product, and they're continuously innovating. And I obviously have a bias towards AWS because I've managed and support the AWS partnership myself. It's good to see that you guys are using it.

One question I have had around the kids solution that you mentioned is that I'm pretty sure you're using AI voice cloning. And just ...

What model are you using? Is it the Whisper model? Or is it ElevenLabs cloning solution that they have? I'm just curious.

Dom Scandinaro:

I think that that might have to be filed away in the proprietary bucket. I will say that we do work really closely. If you look up Cameo Kids, you'll find some launch articles for sure that would talk about this.

We work closely with a company called Veritone, and they're known, pretty well known, for a lot of things in the technology space, but voice is something they're industry leaders in. We work really heavily in partnership with them. They're actually building and running the voice models that we're using and [inaudible 00:31:03] API integrations between our system and theirs as the bookings come in and the videos get generated.

David Joy:

I mean, this is awesome. I mean, thank you so much for breaking down. I'm always excited to talk to people about generative AI, especially since the last one, two years, but these seem very thought out, well thought-provoked ideas that you're trying to implement, and I believe that you might have to start planning for scale and traffic boost right after these things come into production because obviously you're focusing on adding better customer experience, and that's going to support your traffic.

Are you also considering some changes on the platform as you scale and think about these new features that are coming into the platform?

Dom Scandinaro:

Yeah. I mean, I think, just like the AI models where we don't approach scalability with the same hammer for every problem ... We talked about the Fan Clubs and Cameo Calls product before, and that brought with it its own unique scalability challenges. And we approached those with one method. Actually, took many iterations to kind of figure out the right method there. But what helped us overcome that challenge might not help us overcome the next scalability problem. We've just had to look at things uniquely and with a fresh set of eyes for each individual challenge.

I think the beauty of some of this GenAI stuff is some of the things we're doing, like the summarization use cases, are all done off [inaudible 00:32:30], right? We can, on a regular interval process in the background ... New summaries for people's profile pages. New review summaries based on more recent orders that have come in. If you had a worst performance creating Cameo videos a year ago, and you start doing better, we obviously don't want that review summary to live on. We want it to be refreshed with your ... Be more representative of your more recent work on the platform. We're able to do all that offline, and we can even squeeze in human reviews before we end up publishing them to the website as well.

And then for some of this other stuff like the chat-based interactions, it's built on top of Bedrock, same AWS technologies that are going to allow us to be able to scale up and scale down as much as we need to. And ultimately, it's going to be more of a cost-scalability problem, I think, for us at the end of the day than it is a technology one.

David Joy:

Amazing. I mean, that's awesome. Let's switch gears a little bit. And you started as an engineer. Now, leading head of engineering. Now, CTO. Bunch of roles. Tell me a little bit about how the transition has been for you, personally.

Dom Scandinaro:

I mean, it's been ... Leadership is something I've always been kind of drawn to since I was younger. In one of my first jobs, I think I worked as a lifeguard at a swimming pool and then became manager of the swimming pool the next year. And I worked at that same job for a few years in high school and over the summers through college. It's something I've just kind of always been naturally attracted to and I think naturally okay at. I wouldn't say I'm the best leader in the world or anything like that.

But it's something I really feel passionate about as a technologist is just being hands-on. And at different scales, that can mean something very different. When Cameo had almost 400 employees, and we had 100 people on the engineering team that were reporting up through me, being hands-on meant helping review some technical specification docs, helping make decisions on be a decision-maker involved in the decision-making process of using a new framework, or carving out more time for technical debt the next quarter, or what have you.

And as a smaller team, being hands-on means ... The Cameo Kids project we just talked about ... I actually was able to write probably half the code for, myself, because it was something that we were excited about, but it wasn't guaranteed to be this massively successful thing at Cameo. It wasn't easy to justify putting our limited product engineering resources on, but instead something that in 20% of my time I could focus on myself.

I think, regardless of scale, finding the way to stay connected and stay up to date and be refreshing and sharpening your tool set is really important because, at the end of the day, it's really difficult as an engineer. We've all been in this situation as an engineer or any other individual contributor. It's difficult to be asked to do something by a boss that you know has no idea how to do it themselves. They're like, "This is easy. You should be able to do this by the end of the day." And you're like, "You have no idea in the world how this would be done." I don't know how to do everything that our engineering team's doing. That's definitely not true. If that was true, then we would have the wrong team in place because I'm just one person. I can only stay up to speed on so much.

But the team knows that there are areas, especially in the back end and in some of this emerging tech, where I am up to speed. And I'm trying to stay just as up to date as anyone on the team, anyone in the industry, and there are other areas where I don't know my thumb from my forefinger. But just the fact that you're staying involved makes those conversations so much easier when you're asking somebody to put in the extra time. Work extra hard to get something done. They know that you actually know what you're talking about because they worked side by side with you the last time on that project versus you just coming in from nowhere.

David Joy:

I mean, that's awesome. I mean, part of what you're saying is being a leader who's also aware of what's happening gives you that empathy required to communicate ideas. "Hey, deliver this." And you have an empathetic perspective to it, which is very essential. And I think good leaders always inspire but also have a very good sense of reality [inaudible 00:36:41] touch to them, which is awesome. I mean, I'm glad that you're in that role now. And it's really exciting to kind of talk to somebody who has grown from an engineer, too, and has this perspective.

How do you describe the engineering culture at Cameo? In the sense, you're a startup, obviously, and then you will have a startup ecosystem.

What kind of practices do you put in place to move quickly but maintain this high quality that you are with the product that you're building?

Dom Scandinaro:

I mean, I think the perfect thing that happens at Cameo, to give you a peek behind the curtain to help explain and put light on exactly how we work at Cameo within the engineering department and the more broad technical department as a whole, is that we have a practice on the engineering team. I guess we would call it a norm. That every engineer ships code to production on their first day. And that's not a joke. That's literally serious. That's what we do at Cameo. And obviously it's not this massive feature. They're not coming in and redoing the push notification system in one day, but what they are doing is they're joining a team. That team's working on a project, and that project has tens or hundreds of individual pieces.

And when we have somebody starting on a Monday, the week before we're thinking, what piece of this project is small and discreet and simple? Needs done? Doesn't need done this week? We can package it up, and explain it to them when they start on Monday that they'll be able to do start to finish after they've gone through HR training and got their machine set up and all that sort of stuff, but actually be able to write the code, open the pull request, get it peer reviewed, and get it merged and deployed to production on their first day. And we think that is just incredibly important.

There's a lot of pressure, and everyone knows it. It's not that ... Nobody's standing over your shoulder, but the amount of people we've had ... We've had issues. We've had some hiccups come up that, if they've taken it really seriously, they've found a way to make it happen. Other people on the team will drop everything to help them and make sure they're able to get it done. And it sets the precedent that we move quickly. We're not a huge company with a ton of red tape. Anyone is able to do basically anything at this company. They're able to do it quickly, and you're able to get stuff into production really, really fast.

And we take a lot of risks. Cameo is not that serious of a product at the end of the day. If we mess something up, it's really easy to revert things. And just having this practice really pushes us. If we're not getting people's laptops available to them first thing on their first day, they're not going to be able to deploy. If our documentation is not up-to-date, they're not going to be able to get their machine set up. If our IT practices aren't in place, they're not going to have access to the systems they need to be able to get this stuff done. If the product requirements are documented correctly, they're going to be asking too many questions, have too many conversations, to be able to get something done their first day. And ultimately, if we don't have best-in-class continuous integration and continuous delivery practices, we're not going to be comfortable with this person that just started, that hasn't built up trust, shipping something to production on their first day.

But because we have all that, we can be in a situation where we feel not only comfortable with it, but everyone here really, really supports and embodies it. And I think that gives people a sense from day one exactly what the expectations are here in a really positive and encouraging way. And I think, for nine out of 10 people, we are able to determine in the interview process, both on their side and ours, that they're going to be the right person. And some people ... I think this just kind of helps them understand, "Whoa, this is so much different than where I've been before, and I don't like it, and I don't want to be here." And that's fine. We'll help figure it out. But I think, for 90% of people, it just encourages them. "I made the right decision. It's not going to take me two months to get my first thing done. I just did it in the first eight hours that I was working."

David Joy:

Right. I mean, it's also a sense of accomplishment on your first day, and it sets your precedent, right? As what you're going to do. That's awesome. If you have to summarize your experience of Cameo and say, "I learned this one thing building at Cameo," what would that be?

I know it's a loaded-

Dom Scandinaro:

One thing-

David Joy:

... way wide question. I know.

Dom Scandinaro:

I mean, I think this is ... It's kind of a off-the-shelf answer, I almost feel, but it's so true at Cameo is just that you have to be accepting and okay with failure, or else you're just going to build in a silo, so worried about this thing that might happen or this concern the user might have or this annoyance this celebrity might have when they receive a request or booking or whatever. And just have to be okay trying things and knowing that you don't know the answer and being comfortable just building, just shipping, just getting out into production, and learning.

And we're lucky enough to have millions of visitors. And like you mentioned, we've created millions of magical moments. And the time to learn is just so short, especially on our website where 80% of our traffic goes, that we can take risks. We can ship things. They can fail, and we can learn something from them. And more of the stuff we do does not work than what works. But if we're able to fail quickly enough, we can build up a lot of learnings to iterate and ultimately build stuff that our customers want and help our celebrity and creator side of the marketplace make more money, which is why we're all here.

David Joy:

I mean, that's awesome. I mean, at the end of the day, revenue is critical, right? To the business. I'm glad that that is important, and you've learned something like that. And I take something from it myself. I also believe that it's important to learn fast, execute fast, fail fast. All these things work together. That's important.

All right. I know we are coming towards the end of the episode. I mean, this was such a fascinating conversation. I do want to ask you, and this is where we make bold predictions. If I meet you, Dom, next year, part to Summit might or a few years somewhere else, where do you think ... Obviously, generative AI is one of the things that we are super excited about, but tell me some other areas that you feel you're excited where the industry is going and positioning yourself along those trends.

Dom Scandinaro:

I mean, I think there's two things. They're not totally unrelated to generative AI, but both of them kind of existed in our ethos before that became a popular term in the last year and a half. It really starts with our mission statement, and Cameo exists to create the most personalized and authentic fan experiences on Earth. And I say that just to iterate two of those words in there. Personalized. We've talked a ton about that. The chatbot pulling out those details. I think that's clear, but authentic is something that has multiple meetings but is really, really specific. Can be interpreted really specifically towards this technology.

But deepfakes are something we've been talking about since before I started at Cameo. They come up in every interview before GenAI was the thing that people talked about every day. And it's something that we're constantly asked by potential employees. By investors. We've built up a really, really well-known brand at this point. Any portrait orientation video of a celebrity people see online ... They're like, "Oh, that's a Cameo." Sometimes they are. Sometimes they're not. But at this point, we have that brand recognition.

We've been able to build up this stamp of authenticity where, if you see the Cameo watermark on a video, you know it's real. And if you see a celebrity video, and it looks a little weird, and you're not sure, and it doesn't have that watermark, maybe it's not real. And that's kind of the differentiation we've been working in over the years.

But what I'm most interested in is, when we look at deepfakes, what are the 5% of that technology that we can pull back into the product in an authentic way without going too far into the gray area and kind of veering from our mission? And I think where we get most excited about that is in the B2B products. You might have a car dealership that exists in four different cities. And today, if you want to really call out so-and-so's dealership of Pittsburgh and so-and-so's dealership of Cleveland, you have to buy separate videos. If it's $5,000 to get that endorsement once, it's $20,000 to get that endorsement four times. What if there was a way for the celebrity to just say all four cities and it to be stitched in? That's still very authentic.

Going a little further, maybe it's 100 cities. What if there was a way to just, for that one statement in the video, based on similar tech to the Cameo Kids stuff, be able to have a voice model, like a Descript podcast editor-type experience, where change just that one word, and then suddenly you're able to generate that content for maybe 50% of the cost of four videos instead of paying for all four of them. That's where I think we can first start to dip our toes into this area and feel out what the market really thinks is right and is willing to pay for it and what celebrities and creators are comfortable with. I'm really excited to see us get there.

Also, heavily related, when we talked about Cameo Kids, I had mentioned that it's kind of limited today. I'm so excited for a future where, instead of just being able to buy a personalized video for a kid that you know in your life to encourage them to brush their teeth, that instead you could say, "My child ... He's in fifth grade. He's struggling with this specific problem in algebra. And I think that, if his favorite character created a video explaining it to him, it would kind of get him over the hump." The tech is already there to do that. If we give an AI model the exact output we need to go into our video pipeline to generate the video, it'll create a really great script. It'll say, "We have three apples. And if we get two more apples, three plus two is five, and then we have five apples." And then it'll break it down and talk about it even more nuanced. It'll line up the different movements the character's capable of making with the statements.

Getting to the point where this technology is just so heavily adopted and people are so comfortable with it, that we're able to get to a world where, even if it's with generic characters like a dinosaur or something on Cameo and not specific IP, that we're able to create those really personalized learning experiences for children, I think that'd be really cool. And then obviously you can kind of see the potential of how that scales up to things outside of learning and not only being for kids as well.

David Joy:

I mean, that's awesome. I mean, I would love to get learn algebra being taught by Mickey Mouse or something. I mean, this is fascinating. I mean, obviously it's very clear how much you are passionate about this and what you're building. It's very clear in every statement that you've made today.

At the same time, I love the fact that you're thinking about it every day, and it's so clear. One of the things that nobody realizes is, to tweak one word in a voice statement, it's a lot of work. This is a personal story. I had a friend who was getting married. And his brother asked me, "Can you make a voice-loading video?" I've been familiar with ElevenLabs, and I've also used face masking and stuff like that here and there. To change the one word from an actual statement that somebody makes and to make it sound authentic is actually not that easy. But to do that consistently with technology is fascinating.

The first use case you mentioned ... I'm really excited for that because to do that consistently is a challenge and a problem that really excites me to see how you guys will do. Well, that's really awesome, Dom.

Dom Scandinaro:

And I think just a quick add-on to that is you could see the natural evolution of that for one word becomes 10 words, and 10 words becomes the whole video. And I think the most authentic and exciting combination of things that we have there is, if we have an English-speaking talent, and they're able to suddenly scale to speak Spanish or Japanese or what have you, then the marketplace just 3X-ed in viability to entire new markets in a way that's not possible without that technology. That's another thing that I'm really excited about and may be in the more distant future, but the easier stepping stone is being able to start with one word and make sure we nail authenticity first before we continue to march forward any further.

David Joy:

I agree. I agree. Well, all the best with that. I'm super excited to see everything that you guys are going to do. For the listeners who are here, where can they follow you and Cameo and some of the cool stuff that you guys are working on?

Dom Scandinaro:

I mean, I have Twitter. I also have Instagram Threads, since Twitter, I guess, has gone a bit out of style. I don't post there too much, but definitely a good spot to follow. LinkedIn is probably the most relevant place that I actually am and connect with people and am actively messaging and posting new content and that sort of thing. That's a good spot.

Cameo has a couple blogs. We have a consumer blog on BDM that people can check out, but we actually just launched a new blog on our business site. We have a whole B2B product that was related to the car dealership city example we were just talking about. Just posted our first blog post there this week. That's a good place also to stay up to date on the latest from Cameo on that side of the marketplace.

David Joy:

Very cool. Well, thank you once again for coming on, Dom. I really appreciate it. Guys, go check out Cameo. Anything that they're doing is really awesome, especially if you're interested in generative AI. And not just for fan experiences, but also if you're a technologist. You want to learn how Dom and his team are building things. Go. There's much to learn there. Thank you once again for listening in, and I hope you had a great time listening to Dom and the Cameo team. Thank you so much.

Dom Scandinaro:

It was awesome to be here. Thank you, David.