VP of AI at Smartling — Revolutionizing Translation with AI

So in this conversation, we have our guest Olga Beregovaya, who is the VP of AI at Smartling, which is the number one AI translation platform in the market. Olga is probably the most experienced person that we've ever had in the podcast. She's got 20 years in the machine learning AI industry. She comes from a background of linguistics, knowing 4, 5, 6 languages. So it's a super interesting conversation, especially because of her experience and especially because of her role as the VP of AI at Smartling and just the amount of insight she has into the different models, how they work, how translation can actually be improved with the help of AI, like how linguistics actually works into this kind of conversation around machine learning models. So if you're looking to learn a bit more kind of technical details about how AI works or what a VP of AI does and how voice translation is actually processed with AI, then this is a perfect conversation for you. I hope you enjoy. Alrighty, so today we're back with another episode of the AI Report and We're talking to, interviewing someone that's probably got the most amount of experience that I've personally spoke to, that we've spoke to on the podcast of 50-odd episodes. The majority of people we speak to usually have 1 or 2 or 3 years of experience, but I can see your kind of AI experience specifically relates back to 2017. So I want to introduce today's guest. Olga Beregovaya, thank you for coming on.

Okay, thanks so much for having me.

An absolute pleasure. So there's a lot I want to go over. You're obviously VP at Smartling, which is like the number one kind of AI-based, AI-assisted language translation tool. How would you phrase that?

Let me see. Powered maybe, or AI-led since I think probably AI-led is the best one, right?

AI-led.

AI-first. So I'd say AI-led is probably the best way of describing it.

And I really want to come into that as well because I think through doing research on you, it's a very kind of circular route that you've come from those early years of studying linguistics at university and stuff. So I was wondering if we could take things back to then and kind of like your earlier years and your background. I was wondering if you could kind of tell me about your kind of childhood, your teenage years and how that's informed kind of your experience to like where you are today.

Okay. So maybe we can go back to childhood and that would be, I actually was bilingual, bilingual, near-native fluent in English at the age of 3. So it just happened so that I was, I mean, at the age of 3, I was already reading a little bit writing in both, in both Russian and English. So I started with languages fairly early and also went to a school that actually specialized in foreign languages. So that was, I mean, that was pretty much a natural path for me. And maybe other people wanted to be an astronaut or a doctor when they were growing up. I was pretty much, I want to be a linguist. I love languages. I was spending time, I mean, as little of it as was possible in the Soviet Union. I spent as much time as I could with speakers of English language, native speakers of English language. So through teenage years, I actually started taking on interpretation jobs at different cultural events. Because if you think back, that was the onset of perestroika, right? So that's actually foreigners kind of started, started coming into, coming into the Soviet Union or Russia. I mean, whatever it was at the time, still Soviet Union, I think. And then I went to the university for linguistics. I need to be very clear, I have not translated a word in my life. So never studied to be a translator. I mean, I did a couple of courses, mandatory courses, but otherwise I was really mostly studying studying what's, what would be referred to as structural linguistics. So that would be syntax, morphology, phonology, and different aspects of different languages. And I think that very much, since looking back, there's, step back, there's absolutely nothing else I know how to do but languages. Well, maybe also producing music and, but that's, that's a separate story. That kind of informed my future career path. 'Cause once you learn structures of different languages. And once you learn how different language families operate, basically that sets your mind into, I will never speak Swahili, but give me 2 days and I will tell you how Swahili works. So I think that's pretty much, I mean, that was a direct trajectory with zero deviation.

Yeah. I mean, it's certainly something unique and I've not seen from many people before in terms of linguistics. I'm wondering how you went from that kind of real infatuation with linguistics and languages and how they're structured. And I want to kind of take you to that kind of first moment where you heard about AI and how it captured your attention. Can you tell me a little bit about your first experiences with it and when you first thought of it, what you first heard?

I mean, I think we all heard the same thing, right? I mean, we heard the Turing test. Right? I mean, from the first things that we were— I mean, we watched Space Odyssey, we heard about the Babel fish. So I think we all followed, like, in terms of, okay, there's something out there outside of us that can actually think, right? So I think that came pretty early on. Again, when me living in post-Perestroika Russia, foreign films started coming in and foreign linguistic theories and overall foreign information series, meaning foreign sources of information, meaning non-Russian, started coming in. So that's when I started thinking that, okay, hey, as it applies to processing languages, does it all need to be done by humans? Or are there actually other means of processing languages? And that's when I started thinking about what we now at Smartling refer to as language AI. Actually, we've even trademarked the concept because I think we're pretty much thought leaders in language AI. So apart from like all the Everything that you learn from fiction, everything you learn from film, it was also always this thinking, how can we apply any kind of programmatic approach to foreign languages? And that's, I think, where I started thinking about it and I started thinking about, okay, what can we do to process languages? And if you look back at the trajectory of my career, I started building out parsers and dictionaries for rule-based machine translation. And that's something that you can actually describe as very, very early days of language AI, right? If you look back at the history of like companies such as Systran, right? Or like if you look at IBM, for instance, these engines, these machine translation engines that you can describe as early stages of AI, they date back to '50s or '60s, right? When I jumped the AI ship, that was already not new. The concepts already existed. We know how much Department of Defense invested into translation and AI for the lack of options, right? You need to— I come from both sides of the fence, right? I'm Russian and US citizen. So basically both sides kind of had interest in understanding what the other side is doing. So AI in language space was an inevitability, and that's basically when I when I started working on rule-based machine translation, it was already pretty mature.

At the point you entered that kind of space, what was it phrased as? Was it called AI? Was it called machine learning? Was it called natural language processing? What was the kind of verbiage that was used around the time?

I think, I mean, I did some research actually before we chatted. The word AI itself popped up in 1955.

Okay.

But so basically it's been used in different contexts, right? If we look at the movie AI, it's just some entity out there, right, with a mind mind of its own that does not think like us humans do. Now, when AI entered my universe, which is the universe of processing human languages by means of using machines, I would say that probably kicked in much later in the game. I would almost say maybe 10 years ago, if not 7 years ago. It used to be natural language processing, right? You would speak about how machines can process human language. The oldest concept I would say is probably NLP.

Got you. Okay. And then with that, you kind of described that it was like your own interest in this. We started developing these kind of early parsers, the dictionaries and stuff. What was your first official role that was titled around artificial intelligence?

Well, then we, again, we should go back to the use of the word AI. I was basically doing machine translation, which kind of is AI, and my role was described as machine translation innovation and technical leadership, was centered heavily on NLP, which again, if we look into the definition of what AI is, I was— it's a question of a title, but in reality I was doing AI-related applications for pretty much throughout my career. Now when I officially— I don't— I need to go back to my LinkedIn, I think I was officially rebranded as an AI leader in '17. I think 2017. But again, it's a matter of branding. The reality does not change. If we look at the history of natural language processing, it was adding more and more dimensions of going from rules to machine learning, right, to deep learning, and then to what is now described as AI.

AI. And your time across the field, what have been the most transformative developments in the technology that you've personally been subjected to or seen happen in front of your eyes over the years?

Okay, since again, I work in the space of language AI, and we speak about natural language processing, again, most things were rule-based, right, which was transformative in its own right when the Interlingua was invented, when the whole idea of going from one language to to the other language using parser and vocabulary. That was, I would say there were several milestones groundbreaking. So rule-based, all sorts of rule-based NLP, sentiment analysis, machine translation, concept and concept extraction, all of these things were transformative. So that would be one, like applying rules to language. So that's first point of transformation. Huge next leap was when statistical means of processing languages, single language or foreign language, were introduced. So I would say that was the next huge milestone where you suddenly done stitching foreign words together, but you actually apply statistics and you actually apply math to it. So next breakthrough was statistical, everything again, statistical machine translation, statistical, I don't know, semantic analysis, terminology extraction, aligners and everything else. So that was a huge leap forward. And that's when the quality of natural language processing and machine translation specifically went up. And then of course, fast forward the introduction of several generations of neural machine translation, deep learning-based machine translation. That was probably the biggest leap forward when it comes to AI and when it comes to language AI. So I would say that would be the next milestone when we went to neural networks as they apply to languages. And again, most of the NLP directions migrated to deep learning-based applications, which was kind of hard for companies who built their business around like grammatical parsers or rule-based technology, right? Because suddenly you have to, you have to catch. And then of course the latest, which is large language models, right? That's obviously the hugest breakthrough we have seen to date. And I don't know, I mean, obviously, you being a podcast host, I'm not quite sure what you do in your mornings. I read about 40 TL;DRs and like 55 different announcements of what language model came out today. So I don't think, I don't think the field of AI has seen a Wild West like this ever, ever, ever before. And I would call it the biggest breakthrough ever. Which comes with its perks and its challenges.

Out of all those resources you're reading, what are you most excited about at the moment? Because you're so deeply involved with it within the research, within what you do. What are the big things you've kind of got on your radar you're most excited about?

Well, I mean, the whole LLM— I don't want to use the word war. Shall we go with competition? Just the dynamics are extremely exciting because this is the time that things move so fast. Everybody's learning from somebody else's mistakes. So I don't want to name any models. I don't want to put anybody on the spot, but Model A comes out, then Model B comes out to compensate to a great extent for the deficiencies of Model A, right? Like, see, Llama does that. Okay, I'll go into some names. Llama does that, and then next thing we know, that Mixtral says, okay, so this is what, like, I mean, you're hallucinating more, for instance, right? Or you have fewer parameters, then they jump to fewer parameters. Then somebody else says like, hey, large language models are too heavy and computationally expensive. Let's actually try with small language models. So I think just this whole landscape, I don't think any industry have— has ever, ever seen anything like this before. So one thing that really intrigues me most is the dynamics of the landscape where a competitor not only follows their own R&D roadmap but also watches the other competitor and like this constant dance, which I think is super, super fascinating. If we, if we don't go into technology and I'll go into technology in a moment, just the very dynamics of the market and complete inability to follow, like absolute inability to follow up. If one really chose to be educated about all the large language models out there, I wouldn't even know how they would do it. Unless, unless they use the likes of GPT to deliver them the daily summary of what's happening in the world of AI.

Yeah, which is essentially what we try to do as well. So to give you context as well, we run a newsletter, so we are constantly on that news cycle too. I wanted to touch on that point there. You said that the most exciting thing about this is like the competition we're watching this unfold with is Meta obviously released that suite of the new Llama tools recently. And I was wondering if you had any thoughts on open source versus closed source. And how we like continue to develop AI?

I think it's always been, I think it's always been the same. And again, I'm looking back at NLP applications and machine translation applications. I mean, first of all, it's very noble of the FAANG scale companies to open source their models, right? And there's been a lot of discussion. I'm sure you're following that as well as to what incentivizes large corporations to actually open source their models. And to a great extent, if you look at Android, for instance, I mean, it's the community giving back, but then that corporation actually capitalizing on the open source, on the open source input. So first of all, I don't know if I have a very strong opinion about open source versus closed source. It's more what you decide to do as a company. Do you decide to just buy something blackboxed? And deploy something blackboxed, right, which would obviously be an approach if you just integrate closed-source models, right? Or how much effort are you actually willing to put into deploying a foundational open-source model and actually put effort into it and tweak it to your needs? So I think both come with their pros and cons, but I think it's also an eternal dilemma of do you want to deal with black box, or do you actually want to deploy something that you can actually work on? And now also in the world of AI, I think there is so much conversation about ethics of AI, about copywriting, that AI should be accessible to everyone. So I think the whole concept of open source takes a whole different dimension when we talk about AI and large language models, right? Because you really want, you do want to make it accessible and you do want people to be able to fine-tune to their needs. And the community, I think, has never been more inspired inspired, right, to give back and to contribute to those models. So, I mean, these are my immediate thoughts, right? And then, of course, we're all following this debate about OpenAI and how open is OpenAI and who's going to build a real OpenAI model and what's actually OpenAI. So again, but it adds to my previous comment about the world has never operated like this ever, the world of technology. And equally, I think the whole notion of open source is being redefined now. Right. We've always had access to Hugging Face models, right? And which is a fantastic thing, models hosted on Hugging Face. But now I think it takes a whole new dimension technologically and ethically. I hope I answered your question. You absolutely did.

And then you just posed more questions in my head as a result of that. Something you kind of stepped on there. And I just want to talk because we're talking about AI models as it stands. I'd love to kind of explore the underlying technology behind Smartling. And why we actually need it as well. So I think most people's understanding of— when they think of language translation, they just think of Google Translate and they might not understand the full benefits of a platform like Smartling or the technology underlying it. So I was wondering if you could kind of COVID all bases and give us an overview of what that looks like.

I think your information is a little bit dated because these days, if I were talking to you a year ago, I would say yes, the world's perception is What is translation? It's Google Translate. What we're hearing more and more these days is, what is translation? I just plug in ChatGPT API and call it a day. So I think now it's almost like, hey, an LLM would do everything for you, right? It would whatever, it would draw, and it would also shine at translation tasks. And back to why Smartling, more often than not do we hear like, hey, I mean, we love your services, you guys are awesome, but we can just plug in GPT and call it a day. So there is so much more to translation automation than just a standalone API. And if we start unpacking it, first of all, you need to plug all your applications into something unless you have an inspiration to sit there and write connectors to Google Translate and GPT and Anthropic or models or anything in between, unless you feel like doing it yourself. You probably are much better off having an enterprise-scale platform, right, that takes care of all of your integrations for you. Then we're in the joy of— in the joyful land of file types and how many file types can actually equally, whatever the engine you choose, be it Google, be it Microsoft Model, right, be it Amazon Translate, go parse out most complex JSON, right, with tons of nested, tons of nested strings. So that too, I mean, platform like Smartling and Smartling specifically, what we're known for is ability to deal with different file formats. Because once you've broken, like, let me use subtitles as an example. Like subtitles come with timestamps and all sorts of things. Once you haven't parsed the sentence out correctly, imagine what happens if you're trying to apply automated translation to a completely broken string? Like a sentence is suddenly chunked into 3 pieces. How much meaning are you going to get from this translation? Probably negative, right? So we already got connectors, we already got the parsers. Then we also— you said Google Translate. Okay, let's say Google Translate for a second. What if you have access to all the best machine translation engines and all the best, like, state-of-the-art, best-in-class language models in the world and had an enterprise platform that could actually manage it all for you and use machine learning to help you pick and choose which model is best for your task. Think how much thought process and overhead and ROI calculations and is it good, is it bad, and the benchmarking are taken off your hands when somebody else does it all for you. And then I'm going to land at something super, super, super important, which is the quality of translation. Neural machine translation as a concept has only gotten this far. It is fantastic. I know that if you remember 5 years ago, claims were made that the human parity has already been reached, and then there were a lot of counter articles that, hey, the human parity has not been reached, it's your model that was overfitted. So the quality of translation that you can achieve just using either neural machine translation alone or language models alone, it only takes you this far. And now imagine you have a platform that combines the best of both worlds and takes this translation closer and closer to human quality and closer and closer to human parity. So then it's very important issue of human parity and how close to human parity can you get. And there are multiple ways of measuring human parity. You can measure adequacy, fluency, accuracy. You can apply automated metrics such as, for instance, TER, translation error rate, and you can measure how close to human translation machine translation output is. And if you just take one approach alone, large language model only, or machine translation only, it's only going to get you this. For. Large language models dream things up. I'm still, when I talk to LLMs, and I do it daily in Russian, like, holy Lord, I mean, where do they even get those facts, right? But if you can combine the best of both worlds, if you get the fluency of large language models and factual accuracy of machine translation, and yet again, why Smartling put it all into one platform and let them play in accord, you actually get absolutely fascinating results when it comes to how close to human language output it is. So that's my summary of why enterprise-scale, very AI-informed and AI-led platform is of huge value to you, even when you think you can just plug in a single API. Got you.

I mean, that was a very thorough explanation. So Would you see the kind of difference between a good translation and a great translation on parity with human output as the combination of those models there working in parity with each other?

There are 3 parts to it, I would say. There is part 1. If you, again, if you think about what large language models shine at, and I'm sure you've come across it yourself, you get the most fluent human-sounding output, right? This output can be— and it's not only translation, I'm sure you're playing with the likes of DALL-E, right? I mean, and it's often, it's often human-like, right? But it can be completely erroneous. It can be on, it can be off. And I'm reading different statistics about model hallucinations. Some model developers claim that they've reduced hallucinations to only 5% of the output. Or even 3%. Could be true for English. Seriously doubt that that's the case for foreign languages, but LLM output is very fluent, right? And it catches nuances. Like we, for instance, we play a lot. We actually have a couple of patent-pending technologies. You can play with personas and desired formality levels. You can do that with LLM. You cannot do it with machine translation. Now, accuracy-wise, you get your best results with machine translation. Because it does not dream things up. It actually is going to be factually accurate. So from our experience, you get the best if you combine the two. Like, say you preprocessed your text using AI, like simplified your text or whatever, optimized it for translation, right? Then you run machine translation and then you yet again apply both our proprietary, in this case, machine learning technologies and LLM to improve the output. And this takes you to the best place. There is a third piece to it, which is RAG, right, or Retrieval Augmented Generation, where you don't just let these two things hang in isolation from the outside world. And whoever invented RAG, I mean, God, I mean, reinforcement learning in general and RAG, God bless their souls, right? Because the models, like, you know, the biggest problem with GPT was the cutoff on the data. Right, when it— the cutoff on the training data. And now when you've opened horizons to actually referencing external sources and being able to pull the information from external sources, that took it to yet more higher level of accuracy and human parity quality. So I'd say machine translation, um, actually customized machine translation. You can also tweak your machine translation to your tone and voice. So let's go machine translation, let go LLM and let's go RAD technologies. These are the 3 pieces to moving towards human parity.

I mean, I think that's given people a great understanding of how the actual underlying technology works and it's a great explanation as to why you guys are kind of the number 1 AI-enabled translation platform. I'd love to hear a bit more about Smartling and kind of what your mission is, like what's the North Star, what's the goal of what you do?

I would say the North Star is providing human-quality translation through automated means and making global communication completely seamless. If I were like, if I were trying to, and God, I mean, I'm pretty sure our CEO and CMO will hear it. So I better do a good job articulating the mission. But in reality, it is providing human quality translation through automation means and making global communication seamless. So enterprise-scale human quality translation through the use of AI. Now, what I want to say though, and it's very, very important, I'm sure you're familiar with the concept of human in the loop, right? And it equally applies to AI in the language industry or multiple other industries, right? You know that a lot of companies are making a lot of money developing data for models and fact-checking the output of models. So one of the Smartling missions is enabling humans and reinventing how humans do their work. There will always be content that is just perfectly fine to go through a powered translation process and all the steps I just described, and this content will be perfectly okay to be using AI-only features. But if you think about high-liability content, like what if it's heavy machinery instructional manual, right? And the models are still not fully perfect. And say they get it wrong, pull the lever, not pull the lever. Next thing you know, you dropped a construction crane, right? Or again, clinical trials. It could not be any riskier, any more risky to completely trust AI. So one other— I would say our mission is twofold. A, automate as much as possible where it makes sense, but also enable human in the loop to be most productive and reducing the margin and potential for human error as much as we can for tasks where a human is absolutely needed, where human, uh, what is it called and what do search engines call it? Human curation. So enable delivery of human-curated content in the most efficient way. And if I were to sum it up from enterprise perspective, it's really faster, cheaper, and without any kind of quality degradation.

You mentioned a lot of like enterprise clients and stuff, and I'm just kind of wondering, so you've been with Smartling for just over a year, is that correct? Yes. Within that timescale, what are the kind of biggest areas of growth you've seen in terms of your customer base? Is that mainly on the enterprise level, on the individual level? And have you seen any shifting in trends in the last 3 to 6 months?

1000%. I mean, first things first, what I mentioned earlier, this sentiment of just plug in ChatGPT and call it a day and you don't need any— I mean, that's going to take care of all of your global presence. I mean, that was the initial motion on a lot of like in the overall enterprise universe, not to say anything about Smartling customers, but in general, there was this sentiment, hey, hello, silver bullet found, magic wand found. So we had a little bit of that sentiment and we did a lot of work with our customers, talking to our customers and explaining that, hey, everything we just spoke about, right, that just plugging in an AI or large language model API might not be your silver bullet. So we've seen this motion to, can we reinvent cheap and fast? And that motion kind of faded away the more it became apparent that you need so many more pieces to your global presence, which I spoke about already. So that's one. Second, we see motion towards innovation and great push within enterprises to innovate, not just for the sake of innovation, but in terms of reducing the costs without sacrificing quality when it comes to their global presence. Another one is very, very interesting, what you see in digital marketing universe and how maybe dependency on SEO is a little bit giving way to how can we actually operate in the AI-driven search and like next generation. I don't know even if it's— we are in Web 3.0, I think, or are we in Web 4.0, but in the next generation of generative search. So we see a lot of interesting things happening in the universe of digital marketing, but most interestingly, at Smartling, we have this concept of content mix, optimal content mix. And before you saw a lot of translation going to human translation or just more traditional machine translation post-editing, and now we see more and more content being moved either over to neural machine translation and custom engines, or the combination of what I described earlier, a combination of machine translation powered by AI. But it's pretty fascinating how much content is drifting towards full automation. And that's another interesting thing. With the democratization of AI and pretty much a lot of— like, I mean, there is huge access— there are still a lot of accessible ways you can deploy AI in your applications. What we see is convergence between traditional localization and translation departments product departments and marketing departments, because each of them have their own AI initiatives. And it's great to see the cross-pollination of expertise between different departments, just because of the emergence of AI and accessibility and availability of AI. I'll use one example. Think designing a, designing a global interface, like you've designed an interface, right? And it works perfectly fine for English, for your string length in English. And now, hey, you're going to Japan or you're going to Germany. Now we both, you and I, probably know enough about Japanese and German to know that Japan is going to look like this, right? And German is going to look like this. And now with the use of AI in the development process, multilingual development process, you can actually have a very good idea of what your user interface should look like. So one of the trends is we see a lot more globalization, internationalization, and kind of global thinking effort on the part of different stakeholders from different departments.

Hmm. On that note of different stakeholders from different departments, how is running an AI-enabled company or an AI-focused company different from running a traditional company? Obviously, you've been in this AI space for the past 7 years in an official role, but Is there anything that comes to mind? Are you kind of more fluid? It's easier to run? Like, is there anything that comes to mind of how it's different from a traditional organization?

I think we can go back to the beginning of our conversation where I said that the world has never lived in this Wild West environment before. And now I'll go back to you with a question. So if you were running a company that really We are not AI-powered. We are an AI company. And you run this company in the world of omni-changing things. How much easier was life 5 years ago when things were much more static and new approaches and new technologies were not introducing daily? I would say that the most— the biggest challenge and the biggest difference is one has to be agile on steroids. Not just agile, period, but agile on steroids. And there is a very fine balance between frenzy and ability to pivot, right? I mean, either you can just go into the panic, it's like, okay, we need to deploy this model, that model, that model, we need to try this approach, we need to go from whatever, human reinforcement learning to RAG to something else. And I think the biggest difference is how focused on your mission and your vision you need to be now as an AI company and how measured and pragmatic you need to be in your approach on how you want to develop your applications. Because it's so easy these days to fall for the temptation of frenzy, right? That one needs to be in equal parts agile and exceptionally structured and focused. I would say that's, that's the main difference. Like, I don't even know how a waterfall company can survive in the modern time and day. How is it even humanly possible? I also manage the R&D department for our company, and it's also that we cannot pivot from task to task each time new AI approach, a new idea comes up. We really need to make sure we execute, possibly fail fast because R&D is there to fail fast and deliver. So let's say that's, that's probably the main and the huge challenge of staying appraised. I think the fear of missing out has never been bigger in the world, right? And I think the FOMO, how FOMO, right, is probably the biggest factor equally for companies providing AI and companies deploying AI. So that's live in extremely dynamic world. While being super focused on your mission. I think that's the main difference.

How do you make those decisions as kind of new technologies are brought about, new approaches and stuff? What makes you scrap one and put it in the bin and cancel it early? And what makes you take one to the next level? Does it need to reach a certain standard of performance? What does it need to show for you to take that positively and take it seriously?

Well, in every product universe, there is a concept of pitch a ditch, right? So I think the pitch-a-ditch cycle has shrunk significantly. And I mean, first of all, is— first of all, you just learn about the approach and you see how applicable it is. Is it going to improve potentially your translation quality, your relevance of your translation output, or any other linguistic task that you are performing, like terminology extraction? So first, I think by now we know enough to know that something is maybe the next shiny object and we don't need to go there. So that's one. Second one is once you think, okay, this model is awesome, but it's not free. I mean, AI doesn't, AI doesn't grow on trees, right? So then you look at the ROI and then you see whether you're hosting your own models locally or any other machine learning applications, or you decide to tap into the likes of like we are partners with IBM Watson and Amazon Bedrock and Google Vertex. You always look at the ROI. And sometimes you rule out even a super nice model just because it doesn't make any sense to run it on your own GPUs, nor does it make any sense if it can perform the most awesome tasks, but it can be crazy expensive. So next one would be ROI. And then experiments and benchmarking, benchmarking, benchmarking. That's another one. You've adopted something and you run your set of benchmarks. Luckily, we have a lot of data points that we own at Smartling, but we also have a lot of automated methodologies. So it's not that there is a human there looking at which model worked, which model did not work. Quite often, it's a press of a button. So making sure that you keep the benchmark framework always dynamic, I think that's one of the ways, like, you test out the model and you see it's just not going to do. In our universe, there is another dimension, which is multilingual capabilities, right? Some models can be absolutely awesome for English and 5 more languages, but mind you, we as a company do translation automation for over 250 languages. So sometimes the model will just be constrained by the translation directions, the languages that it can service. So usually you have a combination of factors that can help you make the decision for— like Greg, for instance, was a no-brainer for us because we have a lot of legacy translations. We have a lot of dictionaries, glossaries, and style guides. So that one, for instance, was an absolute no-brainer that yes, we need to do it for our translations to be on brand. Actually talking about Smartling mission, helping our clients promoting their brand Globally is another dimension, and each company has their own tone, voice, and style, and you want to make sure that your AI application, your language AI application, actually is able to capture all three and not contaminate your customer's brand. So adaptability, actually adaptability is another factor where you would consider, is this approach good for me? Is this model good for me or not?

I mean, there's a lot in there to unpack, and I think you've given a good breadth of knowledge about what that process looks like. And I'm curious, so obviously you've got a lot of experience and you're very involved in the R&D process. Talk me through what a VP of AI does. What does your day-to-day look like? What's your responsibilities?

What's your role look like? Well, I would say not to overstate the importance of my role in the technology world and in AI-led and AI-developing, AI-leading, AI-transformative company. AI is plugged into every aspect of the company's strategy, right? So which basically means that I'm equally in conversations with our CMO, about how we position ourselves in the market and equally in conversations with our CEO on our next strategic move, equally in conversations with our Chief Revenue Officer as to what is going to speak to our customer base. So I would say there is a lot of executive strategic planning. That's a part of it. The other part of it would be I'm making sure that we drive and my team drives market awareness. We are leaders in the market, no question there. And that's one of the reasons I'm at Smartling now. We are leaders in the market, but the competition is not asleep, right? The competition is also doing a lot of very, very interesting things. So I think one part of my job is making sure that we're very appraised of what our competitors are doing. So that's another one. Now we have a huge, diverse customer base at Smartling, everything from luxury brands to biotech with government in between. Like literally, if you looked, I don't know if you did have a look on our website at our customer base, it could not be any more diverse. So one other thing is a lot of customer interviews, a lot of customer conversations. I make sure that as an AI company, we listen to our customers and our customers' needs. And trust me, the needs of pharma company are dramatically different from the needs of video streaming company when it comes to AI. So internal strategizing, market awareness, a lot of customer conversations, deployments for these customers, because I also manage the deployment department. So whatever the AI application has been deployed, obviously, I'm very much in the trenches into did it or did it not work. And if it did not work, what do we take to product? Then there is a whole aspect of AI features in the product. And there is a super extremely talented lady on my team, shout out to her, uh, who is the head of AI product vision. So that's when we work very closely with product on what features are essential for the market and what features really move the needle. And then last and not least, I'm also very much, uh, involved with our R&D department and, uh, basically making sure that we are moving in the right direction, we don't get trapped into the extremely exciting yet maybe not very productive world of R&D lab, right? It's so easy when you are in R&D to get stuck in the lab and get super excited, right? So also part of my job is making sure that we have very clear KPIs for all of our R&D goals and we deliver towards KPIs and we don't live in R&D bubble, but make sure that we're very, very much aligned with the business. Vision and the value we deliver to our customers. So I, so far, I counted 5 hats. Maybe there are more.

Yeah. So with those 5 hats, it really leads me on to my next and my final question, actually. So you're obviously, you're part of all these different departments, you know everything that's going on. And I realized that through my research, Smartling has been named one of the best places to work in 2024 and has done for many many, many years. What is it about the company culture? What is it about the company? What is it about people that makes it a good place to work?

I mean, first of all, and I loved, absolutely loved, Will Localize, where I worked for 12 years, absolutely loved. And it was extremely painful separation because it was 12 years, for 12 years it was my family and it's a fantastic company and I will always love them and only have good things to say about them. But what attracted me about Smartling is the innovative spirit, is the innovative spirit, no fear of failing fast, pivoting and moving on, and extremely dynamic environment, extremely dynamic development. I mean, I'm absolutely blown away in as how quickly you can tell product and engineering that this feature is essential and how quickly it's actually going to be put into production. So the agility And the innovative spirit of the company is probably one of the, one of the factors. The other factor I would probably mention is just because the company is doing such groundbreaking and transformative things, everybody's extremely inspired. So I would say that, and you only want to be in the inspirational environment, right? You don't want to, you don't want to be in the stiff environment. So I think, I think how inspired people are is another thing that definitely keeps me very happy. Um, shout out to the CEO if he can ever hear this. It's, it's, it's reasonably easy to explain why certain investment into R&D is necessary and how it's going to move the company forward. Once you make a reasonable business case, it is, it is always not as challenging as it could be to get the research funding. And I'm really, really excited about the willingness and vision of the company that really makes big investment into the R&D. So that's very, very dear to my heart. And just like, I don't know, cool people. I always, always look forward to company offsites, company annual trainings, always a pleasure to hang out. So yes, I guess, but the innovative, probably the innovative spirit is one thing that really, really, really keeps me super attached to the company. Super inspired.

Yeah. And it's something I've found kind of common across the interviews I've done with people and that we just have internally as well. Our company is that kind of like that spirit and the speed of innovation is something familiar across the other AI companies that I've spoke to. And it does make you more passionate, makes you more interested in coming to work and being able to actually make a business case for a new development and have that go through without too much bureaucracy really does kind of lift the spirits and make you want to propose new things, want to try new things, etc., etc. So totally understand that, and it's common across the interviews I've done too. I just wanted to kind of finish off and just say thank you so much for coming on, taking the time. I really appreciate it. I know you must be incredibly busy. I was wondering if we could round off and if if you could tell us a little bit about where people can connect with you, get in touch with the Smartling folks and so on.

Okay, so I mean, first things first, Smartling, it's the easiest way to connect with us is just through our LinkedIn page, right, or through the website. We're extremely responsive whenever there is any kind of inbound question or request. So Smartling website would be very easy. I would very much invite you to follow Smartling on LinkedIn. We produce a lot of fantastic content on the subject of AI and language AI and not just us, but we also get a lot of input and we often get our customers to co-present with us. So I'd say high-quality content on LinkedIn for sure. And actually on our webpage, you can also watch a lot of our webinars. We do something called Reality Series where we bust some myths around AI in the language industry. And as far as myself goes, literally just following me and adding me on LinkedIn would be the easiest way to get in touch with me. And I always make sure that I'm responsive and I religiously check my LinkedIn inbox daily and make sure that I talk to people who reach out to me.

Okay, perfect. Thank you so much.