What A Next Generation AI Content Machine Really Looks Like

(0:00) Hi, we're doing something special today. If you tuned in to listen to the usual audio episode, you won't be getting that today because we are for the first time going to be uploading the video onto Spotify. You might want to head over to Spotify or YouTube where it is free as well. We are actually going to be demoing a platform today from the lab. You can obviously listen to it. We're very supportive of that, but there will be a visual component that is slightly more necessary than usual.

(0:26) So please let us know what you think. Max Oshman, CEO of The Lab, which is part of Wellcom. I'm super excited to have you on. This is going to be our first ever demo on this podcast. So I'm really excited about that because you showed it to me already. So I feel like I'm already like sitting on a bit of a secret.

(0:51) But yeah, so very excited to have you on. Thanks for being here. Yeah, thanks so much for having me. Oh, well, it's a real pleasure. You know, we have people across lots of different domains, so I would really appreciate it if you could just explain what the lab does and what its role is inside a content creation business like Wellcom. For, I guess, a bit of historical context, the lab is about 25 years old.

(1:13) We're predominantly based out of New York, but also have offices in Istanbul. And in about a week, we'll be launching our UK office. Congratulations. That's great. Thank you. Yeah.

(1:24) And our business is predominantly three things. First and foremost, we have an engineering-led sort of site platform business where we build a lot of sort of larger scale dot com work. Our second business is a customer acquisition business with a heavy focus on performance creative. And recently, a lot of AI tooling where we've taken really our engineering practice and are sort of performing the like applied AI layer inside of a lot of our clients and inside of Wellcom. And as far as Wellcom is concerned, as you mentioned, they are a global creative production content at scale business. About 500 people all over the world, and they're doing large scale sort of campaign execution work sort of across the globe between the U.S. and the U.K., parts of Asia and Australia.

(2:20) Wow, that's great. Yeah, because I first, well, I mean, I've known of Wellcom for quite some time, but specifically regarding this conversation, I sort of got to know you guys because you guys created from the lab a generative AI digital asset management system, which is fascinating to me because, you know, whereas having come from the advertising background that I do, versioning and, and changing things within, within a DAM system is fascinating. But then the more I got to know about what you guys are doing at the lab is that you seem to have created a whole platform that is a sort of a platform to create platforms. And that's when I really wanted to, got, got very interested in, in what you guys were doing. Cause I'm so keen to, to show people the, the digit, the Gen AI DAM system.

(3:07) But I think that like the, The fact that you've created a node-based digital platform that people can build on top of for whatever bespoke AI integrations they want in their business is really fascinating. Yeah. So over the last year, we've been working on this platform, which we call Wellcom Home. And you could think of a home as sort of like a harness or a scaffold that sort of allows us to build any sort of tooling on top of it. You mentioned the generative DAM tool, which we're going to demo, but we have many workflows, many different kinds of tools.

(3:49) built on top of the platform. And essentially what Home does is it seeks to solve many of the problems more at the applied AI layer, meaning deploying agents inside of the enterprise. The Wellcom is one of those enterprises, but our clients are also those enterprises. And it handles things like data governance and security, auditability, analytics. integration with all of the different LLMs and models, all of the logging. So it's essentially the platform that facilitates all the corporate governance.

(4:19) And then many different tools can get built on top of it. Yeah, that's really fascinating because, I mean, I think that like every time a department head vibe codes a tool, I always think to myself, well, that's great, but where are the guardrails on this thing? And where's your data going? And I think that that's a tremendous thing to be solving right up front. Yeah. I mean, you know, vibe coding is very intoxicating in that it makes you feel like you're getting far very quickly.

(4:49) The reality is that to deploy these things successfully inside of enterprises, there's a lot of work to be done that you're not just going to vibe code. Again, things like data governance and security, things like building data pipelines, things like integrations with existing systems. You know, all of these things have to get built to be real, sustainable, ongoing workflows that you can deploy. I'd love to kind of dive in sort of now into the Gen.A.I. DAM system, because I think that like that will be the most visible example of how the home platform can be utilized to build something that has a lot of use in a production workflow, for example. So coming on now to demo the product is Sam Stroube, the head of AI at Wellcom and the Lab.

(5:38) So welcome, Sam. I'm super excited to have you on to demo our first product on the podcast. Hi, how are you? Very well indeed, thanks. I'm so excited to see this demo. Again, we've never demoed anything on the podcast before.

(5:53) So, you know, you are venturing into uncharted territory, and I really appreciate it. So thanks. Not a problem. Yeah, we have a few different tools that we're really excited about as we've rolled out for specific client use cases. You know, we find that there's a lot of similarities between clients. And, you know, when we're talking about, you know, e-commerce retail at scale or, you know, we have a lot of home good manufacturing, appliances, designer hardware, you know,

(6:25) where materials and finishes and the design of a space is really important. That when you get into content creation and production for those types of industries, especially where it's like premium or luxury and the design is really, you know, it's the lifestyle, it's the inspiration, it's, you know, the environments that carry so much of the implicit brand positioning in how they present themselves that traditional photo shoots just are so complicated. And in the hardware space, you know, whether it's plumbing or, you know, cabinet poles, door handles, when you shoot a room, you need your bathroom, you've designed it around the finishes, your shower head, your shower door pole, your toilet paper roll holder, all like suite together and influenced by the wood that you chose for your vanity. And so you can't go shoot 50 rooms.

(7:33) And so what we find in a lot of these clients is, They rely on shooting kind of their top performing products in studio or, you know, on location. They give away a lot of product to designers, you know, for projects in the hopes that they get assets back. Yeah. But there's a lot of people with the popular stuff. It's very hard to get coverage of a deep catalog kind of.

(8:03) one relationship, one designer at a time. So, you know, some of the stuff that we're doing in that space where we've already been building product configurators and and visualizers using G and a lot of like server side render automation. That tech stack that we built really dovetails nicely into a generative AI world. Where, you know, over the, you know, last year, probably the last six months in particular in the image generative side of things, the quality and the control has come a really long way where, you know, you look at it and you're like, okay, this is... equal to or better than the CG render side of things. Cause you know, we do environment CG, it's, you know, has a cost to it that, you know, it's,

(9:01) not the cost of building out a set or going on location, but it's not economical to scale that across all the styles and interiors. We've really been able to take that structured product data and combine it with the latest image models to be able to create big volumes, cover whole swaths of product catalog at once. Fantastic. Well, we'd love to see it if you'd be willing to share your screen and show it to us. Yeah, not a problem. Max, while we're kind of getting into that, talk a little bit.

(9:34) So what we're about to see is a sort of node-based production generation system that, you know, so all of your, you know, for people who don't know what a dam is, it's a digital asset management. And so that would be... all of the assets that a client has in a product catalog and then all of the environments they want and whatever else, and then they'll have everything saved. What you're adding to that is the ability to have another layer of generative ability on top of the assets that you've already got saved. That's right. And it's not just the assets that we have saved.

(10:09) It's that we've integrated a full CG pipeline into the application. So these aren't even just regular sort of damn assets that you'd get from a photo shoot. All of this is coming from... a CG pipeline that we engineered. So a historic use case for this kind of CG pipeline might be like a .com configurator, right? Like you go and you make a couple of choices, you're designing a couch or a car or something like that and you see those choices render in front of you.

(10:40) What we've done is we've rebuilt that sort of pipeline from first principles for an AI sort of powered world. And so everything you're seeing here is actually live rendering directly from the CG pipeline itself. And you'll see that there's many benefits in terms of output quality using that CG pipeline plus the generative models on top of them. give you a really interesting outcome in terms of scale and in terms of quality that you wouldn't get from, say, like photographing these products as a one off and, you know, sillowing them or something like that. Yeah. And I think that like what what people outside of the domain of advertising sort of.

(11:24) may not realize is for example, like a car company will create a CAD file or a digital file that's a 3D rendering of their car, for example. So that becomes one of the digital assets that comes into a system like this. And then your ability to go, well, great. We have a 3D model of this car. Where do you want to see it? Do you want to see your SUV up a mountain?

(11:44) Do you want to see it taking the kids to soccer practice? Where do you want to see these things and whether you use that as creative development? And now with the generation, the the the Gen A.I. capabilities of a lot of these image models, you can do it at a quality that now you can be spitting out a final ad just out of the digital assets you're already holding. That's right. So, I mean, there's many use cases, excuse me, for these assets.

(12:10) So, the use cases could be the .com configurator. The use cases can be that they're used in advertising. The use cases could be that they power personalization on a website or inside of CRM, like via an API or something like that. Yeah. Use cases could be that these assets get sent to tier three, like like retailers or dealerships in your automotive example. So there's a ton of different use cases for these assets.

(12:39) And really, it's sort of the underlying technology of that configurator, you know, built in, you know, with sort of purpose for the, you know, for the Gendam system and for the AI models that create all of those different use cases. So Sam, what are we looking at here? So this is a bath hardware, towel bars, hooks, you know, toilet paper roll holders. And this is a generative matrix that takes PIM product information It tells our system what combinations exist. So we have a matrix where we have the finish of the color of the product across the top and then combinations of products down the side.

(13:23) In this catalog, it's component based. So the rosette that goes against the wall is customizable. So there's literally tens of thousands of combinations. And so here we're looking at, you know, one combination of Rosette and Talbar. You know, we can come into the catalog and look at another combination. It's a different rosette with the same towel bar.

(13:51) And then, you know, we're kind of looking at that across a full suite of finishes. In this case, all the products come in every finish. So we have a really robust and full grid. But our tool is actually talking to our CG render system to say, this cell represents these two components together. Go request the asset for that. And if it's...

(14:20) been rendered before by somebody else or in a big batch when a product launches, that will load pretty quick. Otherwise, it spins up, you know, our render engine, you know, actually outputs and saves the asset. And now, you know, that You know, our clients kind of use that when they launch finish expansions or add one new component to the system. There's a lot of combinations that come out of that. I think they just did a product launch in the last month that, you know, they added one new component and it was 30,000 new assets.

(14:58) So these tools help us like very cost efficiently produce those assets and fill gaps in the catalog, either fill gaps in the catalog or expand the catalog. So in GeoDentive DAM, we think about the product information and the product attributes as one axi of the content. But then we think about scenes as as kind of a third dimension here. And so what we have here is a CG scene. We have a couple of different things that, you know, different angles that we render in CG for clients. But then.

(15:42) through the structure of generative DAM, we can then define and chain scenes together. In this case, we can use the CG scene as a reference input to an AI scene. So in this case, we have a lifestyle in situ scene set up and we can then batch generate assets that have the product correct reference with a brand on brand, you know, environment that takes all of their finishes into a bathroom where everything is suited together. So, you know, the heart. uh fixtures for uh the the water matches you know the polished nickel for uh everything else and you know everything from the vanities to the you know marbles become uh part of an environment that would be kind of nearly impossible to uh

(16:52) to create these scale of assets in traditional ways, including CG with this amount of variety. And Max, I mean, in terms of like when a client's coming at this, what are they coming to the table with? Are they coming in with just their product catalog? Are they coming in with CAD files? Are they like, what are they bringing to you? And then going, okay, we want this to be put in situ in these ways.

(17:23) Where are you taking them from and where are you taking them to? Yeah, I mean, usually they're just coming to us with CAD data, basically the same stuff that their manufacturing partner or if manufacturing is in-house, their manufacturing teams are using. So they're coming to us with CADs. We're taking those CADs. We are doing a tremendous amount of cleanup. We are building a data pipeline.

(17:47) So we're looking to structure everything. um all of the data not just the assets but all of like the the pym data the product data um because that's really key in terms of the the prompting and uh taking you know taking the client from from a cad to an output like the ones that we're sort of looking at um so really they're coming with cad data we're building the pipelines uh we're doing the sort of structured data exercise and then we're putting them into into our process um When does kind of a creative mind get involved? You know, because obviously I can immediately see how this works for a catalog, for example. Great. Bam.

(18:25) I've got a product catalog. I need all of my stuff in there. But when does it level up or somebody go, you know what? I want a mother walking in and she's with her child and. She rolls the toilet paper off and she's fixing a boo-boo on the head. Where does that come in?

(18:44) And what are the opportunities and limitations of what you've got here? Yeah, that's a great question. I could talk to the process and Sam could maybe touch a little bit on the limitations. But from a process perspective, the creative teams are essentially an operator of our tool set, right? They're one of the users. is providing to them.

(19:08) There's a lot of complexity that's being obfuscated here that otherwise wouldn't be if let's say they were just like opening up nano banana or something like that and trying to prompt, you know, and, and get a similar output for, for, for one, all of the, They know that they're going into a secure environment. They know that anything that gets rendered is going to be product accurate, which is a huge deal, obviously, with product companies. They know that they're going to get matching finishes. They know that they're not going to prompt any product combinations that don't go together because we have all of the PIM data. So there's just a much higher level of confidence in the output.

(19:46) And, of course, as you see here, you could see thousands of images at a time just by sort of setting a scene and picking a product category, you know, that you're interested in whether you're putting it into an ad or designing a landing page or whatever you're doing. So they're coming in here. It makes their life super easy. Of course, And I think that's the benefit of using a node-based system.

(20:07) Like both home and this generative DAM are node-based systems. Can you just speak quickly to anyone who might not know what a node-based system is and why it's the recommended or necessary scaffolding for home? all of the infrastructure that you deliver to a client here. Totally. That's a super critical question, a super important question. And it's probably the biggest architectural decision that we made in building our platform.

(20:33) So if you think about like any... any process, any workflow as a series of steps, right? You are picking a file in a dam, you're opening them, you're downloading it, you're opening it, you're retouching it, you're saving it, you're loading it back into the system, you know, you're doing a prompt, whatever. Those are all steps in a process, right? Hopefully repeatable steps in a process in terms of a production process. workflow.

(20:59) What the node-based system does is essentially takes each of those steps and it sort of isolates them to its sort of atomic level. The smallest possible way that you can sort of define a step is a individual node. So all of these tools, like the one that we just looked at and others that we'll show you today, are a series of nodes in a single workflow. Now, why is that important? So it's important for a couple of key reasons. One, when you break steps down into nodes, instead of just like one, let's say, deterministic flow where it goes from step one to 10 as one thing.

(21:36) In an AI world, it's super important that you work in bite-sized steps because you want to measure, first of all, the success rate at any given step. You need what's called evals. So at every step, we're measuring the output. Two, at every step, you want to be able to sort of trace back. So like if you just build, let's say, a 10-step workflow as one single big step and you don't get the output that you want, how do you know where to QC? So working in terms of nodes lets you look at the step-by-step sort of progression of a job and see the input and output at every single step.

(22:16) So super important for creating what is hopefully predictable, verifiable outcomes at the end of a long workflow. You need to be able to QC at any individual step. So taking those two things together, you want to score your output at every single step so you can improve it. You want to score your output so you can tell the AI system this step, the output of this step is at a sufficient level of quality. You're good to move on to the next one. And then you also need that sort of ability to trace back.

(22:49) And for governance reasons, if you're dealing with, you know, customer data or, you know, you know, high stakes sort of content. You need to know what's going in and out of each step. And then again, from a QC perspective, you want to know what's going in and out of each step. So some of these workflows can get pretty complicated, right? Like imagine a workflow, you know, we're just looking at a generative asset workflow, but imagine if that was connected to, say, like an ad builder, where an ad, you know, excuse me, an asset comes out of this system, you know, an ad gets built, then maybe it gets translated, then maybe it goes through compliance,

(23:24) 50, 60, 70 steps long, so they can get fairly complex. So the node-based architecture allows for all of those things to happen. Yeah, because I think that that's what people who just casually prompt an image in Nano Banana, I think, don't realize. In the world of advertising especially, every image has a thousand different images. you know, different elements to it. You know, for example, the relative niceness of the house that is being photographed, you know, it's like some of these things that we've been looking at here look like they may be a bachelor's house.

(24:00) Other ones look like they may be a family house. You know, the brand values of the company will very much direct what the background looks like. And that's what I think that a lot of people who just kind of prompt in Nano Banana don't realize is the degree of specificity and what each element, whether it's, you know, props in the background or the person you've cast for the foreground or any number of those things, the relative cleanliness of the house, all of those things become specific. part of the brand's statement about who they want to appeal to. And so the degree of specificity they need to have over all of this stuff is incredibly high. And the node based systems kind of really help orient each part of that process.

(24:44) So if you have an environment plus the product plus props plus cast, Each one of those has to be dialed into something very specific, and the node-based system really ring fences that into a place where the brand would be okay with it. Yeah, Orlando, you're raising another really good point. There's sort of two other things happening under the hood. that provides those constraints to the models that again, obfuscates that complexity away from the end user. First and foremost, for every client or every customer on the platform, we have what are called brand knowledge centers, which creates a really contextually rich environment for the AI models to consume.

(25:29) So within a knowledge center, you know, for particular client might be their brand standards. It might be the recent campaign. Interestingly, it might be things that, weren't approved, you know, things that got rejected and the reasons why, right? And all of that stuff is going to the models without the operator needing to know that it's even happening. So again, our, you know, what we're trying to do is create content at scale and obfuscate away, abstract away all the complexity from the user for the creative team or for the producer or for the end customer.

(26:02) Yeah. So all of that stuff is being provided to the model. So we have those context-rich environments. The other thing, which Sam can demo in a minute, is we have a whole review and approval workflow built into the tool. And every time something has a revision or every time something doesn't get approved – That goes back into that contextual environment that then gets provided to every subsequent prompt.

(26:31) So the tools actually get better over time. We're not doing like reinforcement learning or things like that, but we're just providing more rich context to the models over time about what is and isn't going to get approved and why. And again, just think about it in terms of the scale that that unlocks. Yeah. without the operator needing to do any of this stuff from scratch, opening Nano Banana and even knowing that if they did that, it would give them a better output. All of that stuff is getting smarter over time.

(27:00) I think anybody who has prompted an image and gone, no, not like that. No, not like that. Or they've gone, no, no, the background has to be like this. And then it forgets what it's doing in the foreground. And then it does with the doing the foreground. And then the background is forgotten about.

(27:14) Anyone who's done that has fallen victim to a non-node based system. which allows you specificity around the image that allows you to go, no, this is what we've agreed on for the background. And I mean, it really was, you know, Laura's now, I think, have gone the way of the ghost. But there was this this thing over the last couple of years of training a Laura, which is a, you know, a a low weight model that specifically dealt with production design or specifically dealt with product or whatever else. And that's how you got a certain consistency.

(27:45) And what's interesting to me about what you're doing is it seems to me that it makes all in the sense of the world to have a node based system for production executions. What I'm so impressed by is that you've gone, well, why wouldn't we have a node based system that we build everything on top of because any organization? any company has bylaws, they have standards and practices and mission statements and HR departments that have lots of rules written down, like that it seems like actually a node-based, like home, your native node-based platform, I think it's fascinating and wonderful and so smart that you've gone Well, yeah, I know we could vibe code whatever we want for any department, but why wouldn't you build it in a node based system? And why wouldn't you build it from here first? Because all a business is, is rules around any idea.

(28:42) I think that's a really important point. And we've for a year before we kind of undertook this, we were trying to take, you know, tone of voice rules and legal compliance rules and prompt engineer, one big prompt to evaluate campaign copy or what have you. And it's from a context window perspective, it's very difficult to have a model do all of that at once. And you're not gonna get the fine grain feedback and change for each of those rules that exist within an organization.

(29:21) And so one of the other benefits, and Max was talking about to this point a little bit earlier, of the node-based workflow is you're trying to accomplish one thing at a time, and it's easy to then chain those things together so you're not trying to recreate the magic formula each time. But we know that if we have this, you know, the brand knowledge center context for tone of voice, it needs, it doesn't need to be 60,000 words. It can get fairly concise, it can be specific to the channel or the consumer type that we're talking to in this campaign and can get reviewed and scored based on inputs from email open rates and click-through rates and all of that from the performance data. It's so interesting because anybody who has had the experience of having a long chat with an LLM and then you've said something and then a few chats down the line, you're like, no, don't you remember?

(30:31) I told you this already. And we have this standard for LLMs where we're like, I mentioned it one time three days ago and you should be remembering it. And it's just not the way context works. Like, you know, large context, especially if you have a long, long, long chat, it sort of goes in its own direction. And there's a sort of primacy and recency. It'll remember its first prompt and it'll remember its last prompts.

(30:53) But a lot of times in the middle, it sort of starts to deteriorate. And so I think that, like, what's so interesting is that you, the node-based system, you know, sort of is, is the, the infrastructure that holds AI's feet to the fire and goes, no, no, no, this is a rule and you have to operate around this rule. Otherwise, what are we all doing here? Yeah. And Orlando, I think even if, you know, to your, to your point, you know,

(31:18) Even if you know or if you have some people inside of an organization that know how to work directly with the LLMs to get a reasonable output, that's a small percentage of a workforce, right? Like we're at super early days of the diffusion of the technology, right? We're in the first inning of diffusion. Yeah. And, you know, what we're trying to do is provide tools for scale, right? We want anybody to come in.

(31:44) We don't want them to have to think about what they're putting into the context window. We don't even want them to know what a context window is. We don't want them to think about what the right reference images are, what the right prompt is. We just want to give them the result. And to take that a step further, there's users of our application that aren't even humans, right? There's API endpoints that just make requests for assets and the system just gives it the

(32:09) So what we're moving towards is as much automation as possible, as much abstraction from the complexity of working with these systems as possible, verifiable, consistent outcomes as much as we possibly can, and then providing this layer on top of it to let the tools scale. Right. Yeah. We were talking about vibe coding early on and the way that it's sort of intoxicating because it could get you to something sort of 80% finished very, very quickly. But it's sort of that last 20% where so much of the work is. And the reality is –

(32:47) AI, I think, is harder to execute at scale than people admit. And I think if you just look at what's going on in the AI industry with the large model companies basically all launching deployment companies to help diffuse the technology out into the ecosystem, it's to solve some of the challenges that we're talking about. anything that's going on under the hood. They just want the asset, right? They don't want to be thinking about opening up 10 different systems and writing a bunch of different prompts and spending 30 minutes to get something that's kind of good that they can then retouch, right? That's not a great use of time.

(33:34) So that's really what we're looking to solve with our tool set. Yeah, because let's talk about automation a little bit. I mean, it seems like the most ripe area for automation for a tool like this is in versioning, you know, and being able to go, OK, I want this for 10 different markets. I want the big five languages. But I know this market really likes towel racks and buys more towel racks than they do toilet paper holders or whatever else. You know, you've obviously got like a brand knowledge center, you know, that where you've got all of the knowledge of the brand.

(34:05) Does that allow you? Sort of, you know, the most amount of freedom. I mean, the inversioning. Totally. So, I mean, we have ad building tools. We have translation tools.

(34:15) We have legal compliance tools. We have resizing tools. Right. And, you know, I think the example that you just made up. OK, great. Like we have an image.

(34:25) Now the creative person, you know, the creative director designer picks an image that they want to make. the basis for their, you know, you know, social ad or something like that. You know, so sort of what next? So, yeah, again, we have ad builder tools that pull in recent promotional offers from from client APIs or where clients can go in and add, you know, the promos that they want to run. pulls the ad into that environment. We have resizing tools that are built around all of the sort of standard ad sizes for whatever it is, paid social or site placements on a customers.com.

(35:03) We've got translation tools that can deal with the multilingual requirement that you're talking about. And then we've got legal compliance tools that can check it. So again, this is like sort of the benefit of those... of the node architecture is that you can, you know, you can use any of these sort of discrete workflows or discrete tools as one-offs, or you can build really complex, you know, multi-multi-step workflows without needing really any development support at all. These are things that our product managers can go and set up, you know, once they're built. So our focus is thinking about, you know,

(35:40) areas like resizing and translation and compliance, building the supporting nodes, you know, to execute those tasks and then making them available within the system for any kind of workflow, right? The ones that we're talking about and the many that exist that we haven't thought about yet. So we could also think about the nodes in terms of like highly reusable components. And it's these kinds of use cases that that truly amoxed. Yeah. And I think in the versioning, your resizing translation, you know, side of the industry, like Wellcom is that's been a big piece of their business for a long time.

(36:18) And there's a lot of traditional automation, both on the print side, you know, using InDesign and, you know, with, you know, digital tools for, you know, banner ads and homepage heroes. I think where there's more efficiency to gain through the way that our producers and studio designers interface with the tool like Home, but I think where AI is really changing the level of effort for those things are in a little bit more intelligently building the templates and the aspect ratios. We have one client that deals with like very extreme aspect ratio changes from, you know, like very long and skinny stadium graphics and, you know, out of home things to very tall things. And that's that's not one asset.

(37:09) But with the the brand knowledge center and the product imagery and the right environmental creative prompt with that, you're not having to create all those versions of the asset, you can generate them from one master. And so there's a lot of places in our business where automation, the first third was already automated, delivery was already automated, but there was something unique in challenging in the middle that, you know, with building a workflow end to end and using an LLM or a generative model to fill that glue, like we get 80% of it done in automation instead of two smaller pieces before. Max, you had said earlier that there was another part of the platform or a different area you wanted to demo. Orlando, you asked earlier about where the creative comes into this.

(38:13) And so I think that is all kind of woven together here. I think in the context of the generative DAM assets at scale, the creative has already been involved at the point that we're getting here. And that's in defining the look and feel of the brand, the reference imagery, the descriptors that become prompt material and how these scenes get set up. The one that we're looking at here is the high end. This would go into AD or dwell and not look out of place. Sometimes they want an asset for social or for a blog post.

(39:03) Creative can come in and define a more snapshot look where the lighting is not as good. finishes aren't as luxurious, but you still get that same product accuracy. And I think a lot of that is defined based on experimentation and setting up and kind of the more traditional art direction that somebody would work with the mood board. And so Max, I don't know where you kind of wanted to go next, But I think we could go show the product canvas. Yeah.

(39:43) This is a product configurator. You have all of the product data powering what combinations you have, again, connected to the CG environment. And then kind of taking that into a node based environment where you can have a product and and come in and start to prompt, you know, I want this lever in a UGC style bedroom. And then you can open that. And now we have a mid-century modern environment with the lever we want.

(40:20) We can do the same thing where we take two or three products together into a mood board, you know, and it's pulling, you know, the material references, you know, through through the prompt, you know. if we want, you know, product, you know, kind of hero photography, this is all being art directed, what product combinations, you know, what environment, what lensing, you know, angle that the creative wants. And then this output then becomes the foundation for, okay, we feel good about this, you know, do we want to take this to, you know, a larger product catalog? Mm-hmm. So we have a lot of clients that build up their campaigns where they start. Everything starts in the merchandising group.

(41:14) They're choosing what outfits, what looks, you know, need to be merchandised, you know, all the way down to here's, you know, the catalog layout with just, you know, key codes or product SKUs. And so. You know, going, you know, oftentimes on our clients ends going from the merchandising group in to the photo team in the marketing group who then creates a brief for us. you know, chasing the product information and the samples through that whole process. You know, we've kind of tried to provide tools that that create more consistency through through that environment. So this is a kind of a shop builder tool that we have where we can say, you know, this maybe is

(42:11) The image for page two of a mailer within the brand for this client, we have sets. We have for this campaign, we have a studio and a beach location. And for demo purposes, we have Max and I in here. And so we can come in here, create a new asset. We can say we want to be at the beach. We want me in this, and we want this palm shirt and Jersey shorts.

(42:50) And then in the direction notes is where an art director could come in, provide additional references, you know, speak more to the angle, the posing, the prompting. You know, we do use an LLM to evaluate all of the components that have been, you know, combined into the shot and, you know, just have a baseline uh, treatment based on, you know, what, what we know about the, the set and, and the brand. Um, and so here we're gonna, um, dress me in these clothes and put me in a, in a cabana. So, uh, why, why that is, uh, generating, uh, we can go back and maybe look at, uh, one that, uh, we generated earlier.

(43:47) And we, we can actually open this one in, in our review tool where, uh, you know, the node based, uh, you know, structure behind the scenes, uh, is helpful to trace provenance of assets. So, uh, you know, this asset, we know every item that was used, uh, in, in reference to generate it, um, And this is actually a revision we did. You know, the first version had the shirt unbuttoned. We came in here and, you know, annotated, you know, make the shirt fully buttoned and revise with feedback. And so before a human retoucher would even get this, creative can iterate through composition or layout.

(44:36) So in this case, because the shirt you know, is part of the Providence chain. Re-prompting for that includes, you know, all the product correct information. And so we're able to regenerate in that way. Yeah, I mean, listen, this is the dream. This is what everybody wants out of Gen.AI, even when they're just, you know, they're just doing it at home by themselves.

(45:04) They wanna be able to be like, you know, to take specific direction very well and go, just close the shirt. and have it do it. But the fact is, if you did this without a node-based system, it would close the shirt. It would also redo the background. It would also give him an extra toe. It would also, you know.

(45:19) That's right. And I think that like, you know, that's that level of specificity is what everybody needs when it comes to, you know, doing this at scale. And I think that this is, you know, when I was asking before about, you know, where does a creative get involved? This seems to be to be more like obviously they're getting involved in putting the creative, you know, ring fencing around where and how and what, you know, what style they want things to be. A designer can go. I want it in an art deco environment.

(45:50) They can have all those references. But this seems to be where somebody can have a bit more of a play. I mean, especially I would imagine that if you're doing this with. you know, Range Rover's new SUV, you're able to be like, I want it splashing through a mountain thing. The puddle I want, like, I want the door dirty as hell. I want, you know, a kid in the back seat going, this is awesome.

(46:12) You know, like those sorts of things that, that, you know, um, that I think allow this to become a tool of creative playtime a bit more. Automotive creatives have been like the most fun to pressure test these tools with because there is such a level of specificity in the industry for you know, how they show different trim levels of the same car. Oh, we never showed this one off-roading, but this one, you know, that is, you know, okay. So there's a lot, there's a,

(46:48) very deep challenge within that space. And it's been very fun and rewarding to see what ends up working in one space translate to another. We've had a lot of fun with that. One thing, Orlando, also that I want to sort of mention about our approach that I think is maybe a bit different from other agencies is we're not like trying to build the quote unquote operating system or like second brain or something, you know, for how the agency operates. I think there are certain tools that we're building that have a universality to them. But really what we're doing is taking this platform that we've built and we're running a real product management process going sometimes on site into our clients for weeks at a time, deeply understanding how they work, what their needs are, what their guardrails are, what their systems look like, what data needs to flow into our system.

(47:54) And we're building tools for... individual clients. So a lot of what you're seeing, again, of course, there's universality for some of this stuff across category. But much of what we're doing is in that sort of applied AI layer where we're going in through a product management process and we're attacking you know, what we call like narrow and deep parts of the creative production supply chain, where we're solving in many cases, you know, very complex one-off creative or production challenges in a very sort of specific area. And that's really more the methodology.

(48:32) It's not this, you know, build the operating system and everybody's going to wake up in the morning and do all their work through here. It's more how do we solve really complex challenges where there's either a lot of cost or a lot of workflow challenge that we feel like AI can help alleviate, you know, some amount of that. It doesn't even have to be 100% of it, right? Yeah. humans in the loop on many of our tools, right?

(48:59) We might automate 30% of a process or 50% of a process and that's okay, right? Like that might solve a really huge challenge. So that's the methodology here. Yeah, I mean, listen, I think that like there's sort of two classes of product in the market right now or productized service in the market right now. One is a product we already know that needs AI integration and optimization. And the big companies are basically just shoving AI down the throats of a product they've already created and going, look, it has AI now.

(49:29) Where actually what needs to happen is every company that has lots of assets has a DAM system or a media asset management system. And what you're doing is architecting it with AI as the central intelligence to that system. So on the one hand, it looks very similar to most other DAM systems. On the other hand, it operates very, very differently in a much more AI optimized way. So I think that's fascinating. But I would like to talk about what you just said, which is you are going into companies and you're not just selling this system internally.

(50:07) over and over and over again and going, okay, we're a SaaS provider that's just giving you this software over and over again. You are going into companies, interviewing them and discussing with them where they need change or where they have bottlenecks. And then you're creating bespoke software to do that. Talk a little bit about that process. Talk about what frustrations they're coming to you with and also talk about how you unlock those frustrations or where you tend to uncover them. I guess just to take a step back bigger picture, you know, about a year ago, MIT released a study that said something like 95% of AI pilots, you know, fail in the enterprise.

(50:48) And as you sort of dig into that data, there's obviously many kinds of AI pilots, right? So like some of that headline is a little bit misleading. But there's many different kinds of AI pilots. There's just getting a license to like co-pilot or something like that or chat GPT. And then there's doing what we're talking about, which is, you know, integrating a bunch of systems together, you know, trying to solve really complex workflow challenges. with all of the data and governance that sort of surrounds them.

(51:16) And if you dig into that study, what you'll see is actually the most likely to fail projects are in the sort of front of the house part of the business, meaning like marketing, creative, and advertising. And obviously, the more complex the workflows, the more likely they are to fail. And so essentially what we've done is we've dug into all of those workflows, reasons why these sorts of pilots fail. And we're building a product management methodology and process around those and a platform around solving a lot of those areas. So it's a lot of what we've talked about.

(51:55) It's the data pipelines. It's the... you know, training so users can get, you know, a consistent outcome without having to be experts. It's the data governance rules and the access control and, you know, meeting IT requirements and all of the things that we've sort of been, that we've been talking about. And so that's really the basis for Wellcom Home as a platform. That's why we built it the way we did.

(52:21) And then from a product management perspective, What we're doing is we're going in. Well, first of all, you know, again, going back to like the Wellcom Lab sort of dynamic. Wellcom, in many cases, is like doing a lot of these very complex workflows on the creative production side for these clients today. Right. So like these are, in many cases, existing accounts where we're going in.

(52:42) We're very familiar with the accounts. We are already the domain experts. Right. Right. We are already the ones doing this work. We already have an idea of where we can apply the technology.

(52:53) And we're going in and we're doing a few things. One, we're doing all of the process mapping, right, to understand the full process flow, not just our part of the process flow, but the full process flow. How does... And offer get made and find its way onto an asset, into an ad and so on, right? So like the full process, we're understanding all the existing systems. We are doing what's called opportunity sizing.

(53:16) So what I was saying earlier, we don't have to be able to automate 100% of a workflow. It could be a really complex, expensive workflow and optimizing 20% of that workflow might have a huge ROI associated with it. So, you know, we're understanding the problem deeply. We're doing a technical analysis on that on the solution. And then we're sizing up essentially is the bang for the buck there is the ROI there, right? The cost and the complexity of building this thing, you know, what is the payback period going to be?

(53:45) And then we're building it, right? It's a pretty traditional product management process from then on, a pretty traditional engineering process from then on, obviously using two big new things. One, Wellcom Home is the core platform. And then all of the LLMs and generative models and so forth as one of the levers that we can pull to solve a complex workflow challenge. Yeah. So that's really what it looks like.

(54:13) It takes a couple weeks. There's all the rage around forward deployed engineering and stuff like that, all these deployment companies. Microsoft, I think, themselves even just rolled out a deployment company to do these things. And I think that's a lot of what the AI world is going to be. I don't think it's going to be SaaS. I think it's going to be these custom...

(54:33) And I think they're going to be heavily driven by services companies. And I think when you look at the sort of landscape of companies that can fix these problems on behalf of enterprise, you have obviously the model providers who have deployment companies who – what do they know about a retouching workflow, right? You've got the consulting companies. And you've got the agencies and within agencies, there's obviously like places like us. And then there's places that are like new competitors that are, you know, coming out, you know, to build these workflows. But I think for us, what we're bringing to the table is 25 years of actually doing this work, right, for these clients.

(55:16) So we know how to do the engineering and we forward deploy those engineers to the clients and also to our internal teams. So it's, be very familiar scene here where you would walk into a room and there's like two engineers, two CG guys, a producer, a retoucher, you know, and an account person like prototyping something like what we're looking at. And it's like that melding of skill sets and that domain expertise that is allowing us, giving us all the context that we need to produce these kinds of tools. That's what we've found as well is that like, you know, at the end of the day, you know, when we were first building our screenplay management platform, you know, working with engineers who had just never touched a screenplay was just fairly pointless, you know, like it was, you know, and so that's why we work with people who have either been writers or they've been producers or whatever else. And they are engineers who have also dabbled in other things.

(56:08) We're, we're meeting with a, you know, somebody today who, um, is, you know, he took six months off to put on a black box theater production off Broadway. And he's a he's an engineer. That's his that's his, you know, his bread and butter. But he wanted to go do that. And I think that that's just tremendous.

(56:27) And I think that, like, you're right, that that language has to be there of both the tech and the domain specificity, because without that, You're just, you know, you don't know the importance of what you're talking about. And it's not all just a data task. It's an applied data task. A thousand percent. You know, and I think that that's what you've identified here.

(56:47) And that's why I'm so interested, you know, in that part of the home platform that isn't just production workflows. Cause you know, like you said, you guys are part of welcome. You have incredible domain knowledge and production workflows, but you also know about creative companies that run. And I think that that's a very interesting thing to be, to be offering, um, with your platform. Totally.

(57:12) And I mean, again, we're looking at because it's more interesting to look at, you know, we're looking at these sort of generative tools. But I mean, there's tools that are legal compliance tools that run on the same principles, right? The same scoring systems, the same, you know, learning loops. There's writing tools that we've built. So all of these things are possible within performance analysis tools, sales forecasting tools. These are all things that are sort of in the system.

(57:42) Well, I also want to point out that it's not that you guys aren't working with automotive clients. I just couldn't get you guys to present the automotive client work. Yeah, sorry about that. You very, you know, well, listen, no, it's, it speaks to, you know, how responsible you are with your client's data. But, you know, I mean, I think that like, you know, that's, you know, a lot of the stuff that we've seen here today is, is, you know, really impressive. I think that like the size of the task with the stuff that you showed, but I think it does get very, you know, very interesting, very fast, just depending on what brand, what, you know, what you're showing of that brand.

(58:17) But I, I, I really think. That the home system being a node-based system for all of company tooling, I think is something I haven't seen before. And I think is very, very smart. Because I think it just, it puts you in a position where you've got a level of specificity and control around, like you said, compliance, copywriting, you know, compliance. contract assessment and all that kind of stuff that, you know, that I think a lot of people aren't doing. A lot of people are just kind of throwing things into rag systems and, and, you know, building up a vibe coded front end and, and that's good enough for that.

(58:58) Yeah. I think most of what I've seen, you know, out of sort of the agency world is like more than operating system model. I'm not saying this universally for everybody. I don't know what everybody's doing, you know, of course, but a lot of what I just see out in the world is like, You know, we need a performance analysis tool. Well, who at the agency is the best at performance analysis?

(59:18) Oh, it's Sarah. Okay, let's go to Sarah and see what she does really, really well. And let's scale Sarah, right? But let's let anybody do what Sarah does. by, you know, loading some data set in here and going through her process. And it's more that operating system model.

(59:35) And it's less this, I would say, more product driven, narrow and deep, you know, model services model, you know, within obviously, our world, which is creative and production. Listen, that's fascinating. Guys, Max, Sam, thank you guys so much for being on the podcast. Thanks for presenting this. Like I said, our first demo, and I think it went very well. So, you know, I really appreciate you guys bearing with us and trying that out with us.

(1:00:03) But I think it's great what you're doing. And I think that the way you're architecting these solutions is tremendous. And I think, obviously, you're solving a very real – just with the generative DAM system, I think solving very real problems and allowing people to have a lot more control over their assets as they need to repurpose them and move them into a world that – needs content assets absolutely constantly in so many different forms all the time. Cool. Yeah.

(1:00:35) Thank you. Well, we hope so as well. The feedback on the tooling has been great. It's becoming a bigger and bigger part of our business. And I think over the next five years, maybe we'll end up being the core of the business. So thanks for the chance to show it to your audience.

(1:00:51) And we really appreciate it. Well, yeah. And keep, keep us posted as well. Like, you know, if you ever have either a new tool or, you know, or some brand new feature or the automotive client allows you to demo their stuff, give us a call. Like we'll, we'd love to have you back. Thank you.

(1:01:08) We appreciate that. We'd like that too. We love showing it off. Yeah. Well, fantastic. It looks great.

(1:01:12) Thanks guys.