#283 Launching Welo Global with CEO Paul Carr

0:00I'm pretty in touch with what robust ROI financial justifications look like. Honestly, just because of the diversity of use cases within language services, I actually think the ROI for localization spend is really complicated.

0:19SlatorPod Host:Hey everyone, and welcome to another episode of SlatorPod. Today on the podcast is Paul Carr. Paul is the CEO of WeLocalize, language solutions integrator in our terminology and super agency. Yeah, WeLocalize. Hi, Paul. Thanks so much for joining. Thank you so much, Florian. Where are you recorded from today? Orange County. So somewhere in between Los Angeles and San Diego. Yeah, different location for me as well. For those of you watching YouTube, we're up in the mountains here on Easter weekend. So, all right, Paul, let's do a bit of intro first. You did a lot of things prior to Relocalize, prior to becoming Relocalize CEO and consulting financial services.

0:59SlatorPod Host:I saw that you also were in a role with Williams Lee, who was one of my key competitors in my LSI days back in Hong Kong, competing with them. They were also clients, so client slash competitor. So tell me a bit more about the career prior to Relocalize and then, yeah, how this all got started with Relocalize. Basically started my career in strategy. So I was a partner at the Boston Consulting Group, which is a consulting shop. I ran strategy for a period of time at American Express. And then I sort of switched to general management. So I've been running businesses for the last 15 or 20 years, and those have been primarily private equity backed.

1:35So the first one was actually in the legal vertical. It was the largest alternate legal service provider. The next one, the one right prior to We Localised, was in a data business, actually, in institutional investment research. The Williams-Lee experience, I was on the board of Williams-Lee, actually, so through a pretty material transformation that Advent, who were the sponsor at the time, were driving. And so then that brings us up to 2023, early part of 2023, when I joined WeLocalize, and so I'm sort of three years in. And so I'm relatively new to the language industry. I mean, I've been on the buy side, so I have procured language services, but I'd never worked within the industry.

2:21And so I was drawn to basically the international nature. As you can tell, I'm not American. I'm living in America. I'm Australian, but have lived all over the place. I was drawn to the kind of international nature of the industry as well as the fact that it's in the middle of a very interesting period of kind of disruption, which I'm kind of drawn to. So three years into the journey.

2:44SlatorPod Host:Yeah, that's a great time to join three years. That must have been just when the AI revolution kind of kicked off. So what have you seen? What would you say were some of the most significant changes in the past three years since you joined? I joined in the early part of 23. Chat GPT appeared on the scene November of 22. So that was interesting. Interesting timing. Yeah, yeah, yeah. I mean, look, what we've been doing over the last three years is really narrowing our focus. So I'm a strong believer in do fewer things better. So we've narrowed the focus to a few areas where we think we can be successful.

3:25We have been organizing the business around our clients. and so client focus or client centricity has been a very big theme and then driving specialization in sort of within those client orientations you know domain expertise technology value propositions that's been the second focus and so so that's what we've been doing the other thing I would say is we have very materially increased our investment in technology not just software development but you AI and data engineering and so on. So we have been on a pathway of becoming a lot more technology enabled.

4:07SlatorPod Host:Somewhere I read that you said that most of your clients, kind of the revenue now sit outside of the traditional localization functions. Like, tell me more about this. Like, how should we think about it? What client segments are you mostly focusing on? Yeah, so as we've been orienting the business around clients, and there are sort of five sort of segments that we're focused on. And, well, let me just go through them real quick with you. We're focused, as we have traditionally have on the localization department, okay? So localization department of large enterprises. We have a second part of the business, which is focused on, you know, life sciences companies, so pharma research organizations, CROs and medical devices manufacturers, primarily regulated content flows within the drug development process.

5:03Third, we have a part of our business, Park IP, which is focused on legal departments and law firms, primarily the intellectual property and foreign filing. We have another part of our business, which is probably best described as a multilingual marketing agency. They do multilingual SEO, performance, linguistics, and that sort of thing. They do paid media even. They are focused principally on the marketing team, teams of large organizations. And then we have Wheelow Data, the last part, which is really AI training data. They're focused on foundation model builders and labs. And so, you know, as we've kind of oriented the business around these five areas, you know, what we found is over time that the majority of our revenue, two thirds of our revenue, in fact, does not come or no longer comes from the localization department.

6:01So these buyers are content owners, they could be patent attorneys or clinical research managers or, human data training teams and the foundation model builders. And so the kind of the texture of our business is quite different today than it was, I would imagine, a decade ago.

6:21SlatorPod Host:So arguably, We Localized was one of the first true LSIs, like in the sense that we describe it, right? The language solution integrators. Because I remember that when I spoke to Smith, maybe like four or five years ago on a podcast, He also outlined how we localize this tech agnostic, essentially integrating all of the tech. And he described it as best of breed. Is that still the case? Is that still your tech approach? Well, that's good research on your part, pulling up a conversation with Smith five or six years ago. We have traditionally been technology agnostic. And the origins of that was that we're not a TMS, as many of the larger providers also have TMS platforms.

7:03We're not a TMS provider. So we have been historically technology agnostic. We work with practically all of the TMS systems in service of various clients. And when it comes to TMS, we're still not a TMS provider. So we are tech agnostic when it comes to working on various TMS platforms. However, I would say the technology landscape is obviously changing. And we have a strong belief that the agentic layer is going to play a really prominent role in the tech stack in the future. And traditionally, it's been CAT tools and you could say machine translation engines. you know we might get into this in a moment but what we've seen is that agentic systems just do a much better job at processing multilingual content and so we think that's a really important layer and we have also found that unlike you know developing SaaS platforms or you know traditional software you know getting agentic systems right is really requires you know the domain expertise of a practitioner, obviously combined with, you know, data science and technology capabilities.

8:22But, you know, when it comes to agentic systems, we are kind of all, we are not agnostic when it comes to that. This is our Opal system, which we can, you know, talk about, but that's something we've developed and we're really committed to. So when it comes to TMS platforms, yes, we're absolutely agnostic. When it comes to, you know, agentic systems that we think are really important as part of the multilingual workflow. We are not tech agnostic. When it comes to some of the areas outside of the traditional LOC function, you know, who are, where a TMS is not fit for purpose, we've developed a sort of a front end for marketing teams, for instance, and for legal teams.

9:02We're not, those are what we call our studios. We're not tech agnostic for that as well. So I think as the landscape has changed, our posture has changed, you know, with respect to technology and what we really want to own and what we don't want to own.

9:20SlatorPod Host:Has the past year of AI progress made this harder or easier for you or just committing to what you're doing in-house, how you do it in-house? It was pretty clear in that late 2022 chat GPT moment that AI was going to play a prominent role. And obviously innovation has been incredibly rapid. right, in terms of capability. So we started investing heavily in the latter part of 2023. And so we've been on the – we had a hunch that AI in some form agents were going to be important. So we've been investing heavily and in the build process for the last two years. We've been in production in earnest in the last about 12 months with Opal and working out the kinks and what are the boundaries of capability and how does it work in different languages and how does it operate at scale.

10:20And so that's been, I would say this has been a two-year journey, but we're, and in that time, we've been building a lot of capability in terms of data science and machine learning capabilities. And so we've been, you know, we're pretty committed to that.

10:35SlatorPod Host:All right. So when we schedule this podcast, sometimes things happen in between. Queen, so now you've just launched or announced WeLo Global. And I was trying to unpack it before the podcast, but hear it from you. Tell us, what is WeLo Global? As I mentioned, in the last couple of years, we've been executing a strategy of client focus on the one hand and specialization on the other hand, and organizing the business around client segments. And also, as I mentioned, today, two-thirds of our business does not come from the localization department. It comes from outside the localization department.

11:11And that's problematic when the word localization is actually embedded in your corporate logo, right? We localize. And the reality is that for folks outside of localization, the term local of the local department, localization doesn't mean that much. So what we're doing is we felt that we needed to kind of catch our branding up with our strategy and the realities of our business today. So we're adopting Wheelock Global as our new corporate logo. So it's replacing Wheelocalize. And Wheelock Global, think of that as an umbrella that sits above five kind of client brands that face off against very specific client segments.

11:56They're solving specific problems. They're building specialized solutions. And that's Wheelocalize against the localization department, Adapt, our marketing agency. wheelo life sciences which is a new brand we're also launching this week which is a brand you know uniquely facing off against pharmaceutical companies and contract research organizations and device medical device manufacturers park ip which is our legal brand and wheelo data which is focused on foundation model builders and uh and labs and so each so think of wheelo global as an umbrella right it has a connective tissue which is our culture and you know our sort of core capabilities in multilingual generally but our client brands are really the brands that are facing off against the market and so each of those brands is overseen by a general manager they have their own client you know dedicated client delivery or operations teams go to market marketing product and technology teams, all focused on solving that client segment's problems.

13:02The brand evolution really is evolving our branding to be more reflective of our business today and our strategy and where we're headed.

13:12SlatorPod Host:Yeah, sounds really familiar because we just did that a day ago for those who have checked out our website. So anyway, we didn't change the name. I'm going to check it out. Yeah, we changed everything else. So what was the thinking behind keeping the Park IP and the adapt and not renaming it as WeLo, I don't know, Adapt or WeLo IP. Like, I'm sure you had a lot of discussions around that. Was the brand name too strong? Was it something you didn't want to give up on? As we were going through the thought process, we knew that WeLocalize no longer served us from a corporate standpoint. But in terms of what came below, we did a lot of work.

13:46We did a lot of surveying of clients out in the market. We did a lot of SEO analysis, like who types what into a Google search bar and what are they looking for? And it turns out the Park IP brand is one of our strongest performing brands. And you may know Park IP, we acquired Park IP a decade ago. And we've been playing around with, well, is it Wheelo Legal or Park IP, we localize. And we decided, you know, the brand awareness is so strong, the team decided we're just going to go back to Park IP and not, you know, and that's our client facing brand. So we stuck with it. Got it. Adapt, similar story or?

14:32Adapt, similar story, yeah. I mean, Adapt is, yeah, yeah, yeah. So Adapt, again, had really strong market awareness among marketing teams. And so, you know, we stuck with Adapt.

14:44SlatorPod Host:Willow Life Sciences, that's an interesting one. So this is, again, it's one of the five sub brands. So tell me a bit more about generally working in regulated industries. Like what are some of the tailwinds, headwinds there? How is AI and language AI perceived in life sciences? You know, very high cost of failure, obviously. And, you know, people are probably a little more cautious to embrace, but no, it's been three years. So I'm sure it's happening there as well. It depends where, which workflow. So when it comes to, you know, a lot of our business in that segment is in the, you know, the confines of a clinical trial, right?

15:24So really in the core of the drug development process. um there where the the workflows are a reason are regulated the translated the patient feedback and so on on on drugs um has a very regulated process around it that content goes to regulators in local jurisdictions um and if there are errors or there's sloppiness around that process it causes real problems reputationally for the sponsor, for the pharma company. And second of all, it can slow down. The more you start to get iteration loops in that process, it can slow down the clinical trial. And a clinical trial is an incredibly expensive process.

16:14And the more you slow it down, the more you reduce the time at which a drug is going to be on patent. So it's a big deal. I mean, each day is worth millions of dollars to a pharma company. And so if there is, that is by far the most important component of the value proposition. It's not about cost. To a clinical case manager, the cost of translation is kind of a rounding error in the budget of a clinical trial. What really upsets a clinical case manager is if there are escalations with a regulatory body in a far-flung geography, or it starts to slow the clinical trial down. So, you know, I would say the embracing of AI is tentative and much less around cost and more around how do you ensure absolute consistency of precision and quality so you don't bump up against kind of slowing things down or, you know, leading to escalations.

17:12SlatorPod Host:And there you work mostly with the CROs or also sometimes directly with the sponsors if they run the study? Yeah, both. CROs, because obviously there's a very large part of the drug development market that is outsourced to contract research organizations. But by the same token, there are also sponsors, right, pharma companies who manage the clinical trials themselves. So we work with both. You also spoke about WeLo Data, your AI training data business. I think you've been involved in this for quite some time. You know, we just published a big flagship report of it, the 160 pages data for AI report.

17:49SlatorPod Host:I hope you had a chance to see it. If you don't, you should. So this is, we see it as a key growth area for LSIs. What are your thoughts on that? I mean, how do you see the business perform? I'm really glad you covered, I have read the report, 160 pages. And I'm really glad that you covered it because it is a natural adjacency to language services. And frankly, that space, it doesn't get much coverage by anyone. So I thought that that was a really good report. Look, I think we had been involved in some capacity for a number of years in AI training data, but we were managing it as a bit of a side project, right?

18:32And it was bumping along. we got when we went through our review of the business in early part of 2023 we are really excited about the space right it's very dynamic it's growing um and it's sort of a core part of the infrastructure of the you know the ai revolution so that was the part of the business we kind of separated out and gave more autonomy to first and that that was in the back end of 2023 What I can tell you is that it is a growth opportunity for LSIs, but it is a very different business. I mean, when you really kind of get into the actual operating model and the technology requirements, it is quite different.

19:19The workflows are different than traditional localization. The talent pools are different. What the talent's doing is different. The buyers are different. What they're buying is different. The other thing I would say is that because it's kind of under the umbrella of all things AI, it's a very competitive space and it's getting more competitive all the time. So you have the sort of Silicon Valley natives like Scale AI and Surge and Turing and so on. But then there's recruitment platforms who have entered the space like Mercore and Handshake. There's kind of companies with vast gig worker networks like Uber and DoorDash, traditional BPOs and systems integration companies like Telus and Taskus.

20:07So it is a very busy space. it's part of the reason why you know with with separating the business out wheeler data um you know our team there today is it looks nothing like it did a couple of years ago you know we have a very talented general manager siobhan hannah whom i think you may have met she ran linebridge ai prior to its acquisition by telus and ran the data business at telus for or four years. But most of that team are kind of X scale, invisible app and, um, or from the buy side, right. That the human data, you know, teams at Google meta and Amazon. So, so I think there's a fair amount of building for an LSI to be successful.

20:52I think they probably have less than 50 % of what they need. Um, so there's a fair amount of sort of tailoring and building that needs to happen to be really successful in that, in the sort of training data space.

21:05SlatorPod Host:So they don't really have a structural advantage. I mean, they do have maybe, yeah, you said 50%, right? So a bit, but then they need to make a very conscious decision to go after this market. I think it would be tough to call it a structural advantage. I think they've got a head start specifically around multilingual data sets. You know, so look, the Silicon Valley natives, Scale and Surge and Mercor and Telus, or Telus Digital to an extent, they have not built out right full multilingual data capabilities because it's hard and you know lsis have been doing that for a long time and you know it's it's not easy to do that so that that is where you know i would say lsis have a head start but even then there's a fair amount of adaptation that needs to go on and like i said the talent profiles are different you know the evaluation how you evaluate workers is quite different because they're doing a different job and what we found is just even the volume of applicants in the sourcing funnel is kind of an order of magnitude greater so yeah there's a fair amount of kind of adaptation that needs to go needs to go on even with a starting point that's you know a step ahead of the kind of the other players the silicon valley natives if there's one more follow-up on that like do is there still a lot of kind of startup activity or some of these, I mean, scale has been around for a while, right?

22:34Yeah, scale kind of, you could say founded the space, probably not founded the space, but was early. There's still a bunch of startup activity. It's incredible. I mean, there are startups all the time. They tend to be people out of that Silicon Valley ecosystem. Then maybe they're coming out of one of the sort of foundation model builder labs and they think they can do something better and have a better value proposition. So yeah, it's pretty busy in terms of startup activity as well as other companies in adjacencies who are also trying to enter and build those capabilities as well.

23:16SlatorPod Host:But the secular need for that type of data among the hyperscalers, the frontier labs, that's here to stay, right? Because I've spoken to people who said, you know, maybe we're kind of doing another round here a couple of years. And then, I don't know, you're all going to like everyone's just going to use synthetic data. But I guess there's going to be new ways of leading these companies and these capabilities. We think it's there to stay. It's certainly not stopping growing. I mean, the demand from the model builders is not slowing down. Now, the nature of the data requirement is different. So I think that several years ago, it was more general data that was required for training models.

24:02It's now, there's a lot more specialist data. So, I mean, there's PhDs who are generating data for training models in certain specialties. Multilingual is a really rapidly growing area because, as you would know, the kind of pre-training data set, the kind of common crawl, all of the content on the internet is very oriented toward English. So in the work we've done, but many others have done, the models just behave far less well outside of English and far less well when it gets to low resource languages. The safety parameters work far less well, right? It's much easier to jail jailbreaker model in a low resource language than it is in English.

24:49So the model builders are, you know, are attentive to this. And so, you know, multilingual is a big deal. Multimodal is a big deal. So visual or auditory data is people are building the capabilities for chat bots and, you know, agents that work in a customer service environment. Robotics is only just really getting off the, that requires a massive amount of data. So we don't see the need for human generated data slowing down. Now, there is also synthetic data that's being used in the process, but the entire space is growing both on the synthetic side as well as the human side.

25:28SlatorPod Host:Let's talk a bit about Opal, the system you mentioned, the integration system. So yeah, tell us more about it. I guess WeLo data is probably not on it. You're not using opal for the data business so that's more that the localization side it's more the localization side so basically opal is in very simple terms it's an agentic system that consists of machine translation but llms um to do translation work right so the way to think about it is it is a much better version of a machine translation engine in fact it's twice as good in all of the testing we've done, it delivers content that is twice as good from a quality standpoint, however you choose to measure it.

26:14Now, like I said, this has been an area of investment for us for the last couple of years. And we've been in production for 12 months. We've done a lot of testing, a lot of side-by-side comparisons, and it just performs much, much better. And so one of the questions well why does it perform better and part of it is that in putting an agentic system together it requires a really robust harness that allows this the models to be trained on an enterprise's kind of unique content and style including language pairs and glossaries and style guides. Secondly, these agents, given their context window, have a much broader context than an empty engine.

27:04So they tend to do better on fluency, which is a challenge with empty engines. But third, there's a fair amount of science in constructing this harness that I was mentioning around the algorithms. And that involves RAG, right, and database architectures, how content is actually ingested and the feedback loops to make the whole system kind of continue learning. So, look, that's what Opal is. The final point I would make is that, you know, The way we're going to market is we're taking a distinctly open architecture kind of interoperable approach. And so what I mean by that is that we have well over a dozen relocalized customers using Opal today where they're paying for both the technology and the servicing.