#278 Bluente CEO on Solving Tough Last-Mile Problems in AI Document Translation

0:00When we set out to build our AI-powered document translation platform, we ensured that it covers the full end-to-end workflow from not just the translation of the content, but also the entire formatting of the file.

0:18SlatorPod Host:Hey everyone, and welcome to another episode of SlatorPod. Today, we welcome Daphne Tay. So she's the CEO and co-founder of language technology platform Bluente. Hi, Daphne. Thanks so much for joining today. Hi, Florian. Thanks so much for having me. Really excited to jump into the discussion. So I assume that you're recording this from Singapore today. Is that right? That's correct. All right. One of my favorite places in the world, Singapore, you know, moved there exactly 20 years ago. So, you know, amazing place. Come back again soon. I will, actually. I will. I'll let you know after the podcast.

0:54SlatorPod Host:So, cool. All right. So you're recording this from Singapore. So first, tell us a bit more about kind of your background, your journey here, and maybe also start with a quick elevator pitch around Bluente for those who don't know the company. So first, introduce us to the company in a second, and then just tell us a bit more about your background. So we set the scene here. Sounds good. Hi, everyone. My name is Daphne. I'm the founder here at Bluente. The name Bluente stands for Business Fluent, and we are an AI-powered document translation company. What we do is if you have any PDF files, PowerPoint, Excel files, Adobe Suite files, put them through two minutes or less.

1:35We translate it in 120 different languages, but we keep the format the same. So we're really excited about what we're building. And maybe dialing back to the background of how Bluens started as a company, we actually started as a language learning platform, very much similar to an app that a lot of people might be familiar with, Duolingo. but tailored for business professionals like lawyers, bankers and if they're looking to learn, say, legal terms or financial terms in a different language, they can come to Bluend for that. Unfortunately, that part of the business didn't quite work out but a close adjacency is the translation which is language related and we see a lot of people struggle with not just the translation of the text but also the full document translation.

2:26And that's why when we set out to build our AI-powered document translation platform, we ensured that it covers the full end-to-end workflow from not just the translation of the content, but also the entire formatting of the file. I'll pause there and open for more questions. All right.

2:47SlatorPod Host:I want to go into why this particular problem, I have a suspicion why you kind of came across this particular problem because you were with Bain before, right? That's correct. You were a consultant at Bain. So then let's go back to the origins of this particular kind of problem, the formatting of documents, right? Which at Bain, you must have faced quite often with all the pitch books and the presentations and things like that. So was that maybe where the idea started? Yeah, exactly. I think there were a couple of issues that were faced during my time at Bain as a consultant. The most memorable one was at the end, towards the later part of my career when I was about to leave, we had to submit a pitch presentation to a conglomerate in one of the Southeast Asian countries.

3:41And this pitch deck was, I think, 600 pages long and it was all in English. and it had to be translated to localised because the company required a local version of that deck. And it was, I think, one or two days before the deadline and it's just humanly impossible to translate 600 pages and for one person to do it. And I remember at that point in time, they essentially had one of the subsidiary teams to curb all hands on deck to translate it and make sure that it's really localized so that when the client reads it, the prospective client reads it, it reads fine. So that was one of the instances where you're like, okay, you can't just ensure that the text is translated.

4:28You have to ensure that the text is translated in the context, in the format that can be delivered very effectively. And today, if Bluant existed back then and you put in a 600-page document into our platform, that would maybe take like half an hour or so to run the entire file format. So that was one instance. And then, of course, there are other instances where you are dealing with Excel files. We are dealing with financial statements in different languages. And I remember back then, we received an Excel file and it was localized in, say, Indonesia or Thai. And then you needed to understand what exactly was those line items.

5:07You kind of copy and pasted each one back then. So, I mean, technology is moving so fast and I'm glad that, you know, we're building something that it's really solving such a big problem that we hear from a lot of our clients.

5:20SlatorPod Host:Yeah, I think people might underestimate the size of this problem. In fact, that was during my days in Asia, this was the key problem because we worked a lot for the investment bank and apartments of the big investment banks, right? And so we worked a lot on Hong Kong IPOs, which involve these types of pitch books, right? So, you pitch the big Laoban somewhere, the big boss somewhere in mainland and that person isn't going to read a pitch book by an international investment bank in English. So, you need to fully kind of translate it into Chinese. And so, deadlines are incredibly tight because the investment bankers, they take a lot of time to work on the original book, but then the localization, the translation is a complete afterthought and then you'd have like a couple of days maybe if you're lucky.

6:08SlatorPod Host:Yeah, I want to delve into that. Maybe in a second, because this is, I think, an underappreciated problem in the world of localization. And so let's circle back to that first. You did mention Indonesia, Thailand. But also tell us a bit more about, let's start with Singapore being obviously the hub in that region, a key business hub. So what are your impressions or what's the state of the startup environment in Singapore and then further in Southeast Asia at the moment? So let's just take a broad lens and then we go back to Blue and your idea. I think the startup environment, I would say, I mean, we can look at it in different angles from a funding angle.

6:54I think in recent times, that's probably not as vibrant as in the other side of the in Europe or in North America. We're not seeing so much activity here. I think people are much more conservative because of how the market has been shaped. But I think on the other side, people, a lot of entrepreneurs are emerging. I think with different AI tools that are coming up, the barrier to create a company is lowered. So we do see a lot of startup activity in that sense. And we are starting to see quite a number of companies setting up their headquarters here. Especially Singapore being in Asia, we do see quite a number of Chinese entrepreneurs moving here.

7:47I think the most prominent one recently is Manus AI moving from China here to Singapore. And then, you know, then they got acquired by Meta recently. So we do see a lot of that cross-border movement and Singapore being that space for a lot of prominent ex-founders, especially from Asia, to choose that to be the country of place to set up. and as a result of that we're seeing a lot of like you know movements in terms of founder community and for example open ai has also set up their asia headquarters here in singapore so that has also helped to you know really make the entire startup ecosystem much more vibrant because when you have different you know the big boys here the founders and more and more founders and then that creates like a much more vibrant ecosystem.

8:41So from a building perspective, I think that has been really nice.

8:45SlatorPod Host:And specifically on the language part, the language AI part, I mean, Singapore being a multi, you know, like multilingual place, like one of the most obvious multilingual places in the world almost, right? If you go there, it's incredible the diversity there. Is there any specific focus on that or not so much? just this general AI and general startup kind of buzz at the moment in the city? I think from our product perspective, given where we are located, we have spent a considerable amount on Asian languages. So your Chinese, Japanese, Korean, but also more importantly, the Southeast Asian languages.

9:24So for example, your Thai, Vietnamese, Indonesian, Malay. I think those are languages that in certain instances might be, you know, deprioritize in other markets. But I think given where we are, we do see quite a number of requests from our clients requesting for Thai translations or Vietnamese translations and things like that. So building the product and ensuring that linguistic accuracy, we have spent quite a bit of time, not just on the major business languages, European languages and the likes, but also Southeast Asian languages as a core part of our coverage.

10:00SlatorPod Host:All right, let's go back to the product. and kind of, you know, one of the key USPs, the formatting, right? I think when we did our research, once you described formatting, translations last mile rather than the language. So tell us a bit more about that. So specifically, what are the key challenges in formatting when it comes to the translation part? So many. But I think the key thing is, you know, there are so many different file formats that are being created. Like it could be, you know, if we are looking at, say, PDF as an example, it could, you know, to us, it could just be, okay, it's just a PDF file.

10:43But actually, PDF is like a readable type document, right? So in that instance, what that means is it could come from all different, you know, sources. For example, I take a photo and it becomes a PDF file. It could be like a Canva converted to a PDF file. or it could be like your Microsoft Suite and Excel or Word that's converted to a PDF file. And then you have more complicated graphics like your AutoCAD drawings and stuff like that that becomes a PDF file. And if you dive deep into it, the reason why I'm kind of like sharing the spectrum of the type of files is because a lot of times when we are dealing with these file types, it's not just solving for like one type of file there are like many different subcategories of the file so as an example scan file types is one that you know our platform deals really really well with and when it it might just be then then the question is how you detect like a a word or phrase in a particular language within that scan file type right and then when it comes to you have two columns of that scan file, then how do you determine that this is actually two columns, especially if they're very close to each other?

11:57How do you determine what's that right spacing between the columns to say that, okay, these are two separate columns versus one so that when the machine is reading it, it can read it right that, you know, these are two separate columns. So we dive very deep into each of the different file types. But I think where we have started from as a company was to focus on the legal vertical, then now we are looking at other verticals as well. So within the legal vertical, like what are the different use cases that they will have, what are the different types of documents that they'll deal with? And the interesting thing is because within the legal vertical, they have so many, they service clients from all different types of industries.

12:36And as a result of that, that helps us to build up a strong repository of the types of documents that we need to deal with.

12:44SlatorPod Host:Makes a lot of sense. So you mentioned legal, and this is deliberately targeted initially instead of going broad, so you're saying legal but also financial. Have you found traction in the financial space as well with all the IBDs, investment bank departments? These are the two key ones that we target, but I was going to share a funny story earlier when you brought up about the investment banking teams and late nights and translation being an afterthought. I think this was a team in Hong Kong, and we saw like I think it was 4 or 5 a.m in the morning they started to like dump in 400 pages of files into and then they were just like topping up credits non-stop to our platform and we're like okay you know given that they're from the IBD team it's likely they just completed the deck you know they rushed through the night and they just needed something that can very quickly help them to translate it while keeping the format the same so that they don't have to spend the next I don't five, 10 hours dealing with both the translation and the formatting at the same time.

13:50SlatorPod Host:Incredible, incredible. So this is, I mean, this brings up so many questions in my mind because it was literally, we used to be on the receiving end of the desperate investment bankers dumping a bunch of files. Like I would get like, you know, millions of emails at 5 a.m. in the morning and they would ask us to, you know, yeah, do the translation, but then also formatting. At the time, they had entire internal teams by outsourcing providers that were only dealing with the formatting, like the figure calling and things like that. And the translation was kind of even a separate part. So in your case, is it fully automated?

14:29SlatorPod Host:So the bankers would go there, they would upload all the content, they would buy the credits, they would run it through. And then they would do kind of the final check in the figure calling, et cetera, on their end? or would you have some team or some human experts that would help with that? I think it really depends on the setup of the organization. Sometimes they do have internal reviewers and they can actually just do the... I mean, our system right now, we can automate the full end-to-end, right? So if you dump in, let's say, a PowerPoint presentation, you get that back in a PowerPoint presentation in whichever language that you like.

15:05But sometimes they still need like a human in the loop to review that. So we do offer that service layer as well, if required. So it really depends on the organization and the firm setup and the nature of the translations that are required.

15:23SlatorPod Host:I'm not sure how much you can disclose around that, but confidentiality, obviously a key issue. So we used to have an actual dedicated line with some of these IBDs, and if it got super confidential, we had to send translators over to the office of the big investment banks. Like when they just upload a bunch of files, do you have some type of... I mean, what's the confidentiality protocol look like with some of these banks? We have a zero data retention policy. So in general, we don't store any of the files on our server beyond 24 hours. With the LLMs that we work with on the translation layer, it's deleted immediately upon processing.

16:01So that helps to ease the first layer of security. Then the other part of it is we have, you know, most recently gone through the ISO certification just to make sure that in more broadly speaking, we are abiding by the security standards. So we're ISO 27001 certified. We encrypt all data both at rest and in transit. So these kind of standards have been pretty much acceptable to our clients and they are quite happy with that.

16:34SlatorPod Host:How do you generate these types of leads? I mean, is it just like word of mouth and like some junior banker knows the platform and then says, hey, hey, hey, I have an idea. Like we need to finish this by tomorrow morning. Or is it more kind of more structured where you have like actual campaigns and things like that? I think it's a mix. To be honest with you, like currently as we speak right now, we are also running different experiments on how best we can reach this group of people. we a lot of the when we ask our clients like okay how do you hear about us a lot of them hear about us interestingly through AI if not through Google search because I think it's yeah because with translations and I'm sure you know like it's a very time triggered activity it's something where you know you don't think about it until like it's like oh I need it now and then you're like, oh crap, where can I find it?

17:32So we also try to reach people at the time where they need it. And that's an internal experiment that we're running right now so that we can reach people at the right time. And of course, platforms like this that you have created also helps to bring to life platforms like ours and bring that to the attention when the time comes.

17:54SlatorPod Host:Yeah, we'll make sure to get a good write-up for this podcast. So, the Googling or chat GPT-ing banker finds it when they need it. So, let's talk a bit about language coverage and the challenges. So, we spoke about the formatting, which obviously is super technical and very, very difficult, but what about the language part? Maybe starting with Chinese and then like moving to other Asian scripts, maybe Thai and just other like challenges specifically to the Asia market. I think Chinese as a language, it's probably, I would say, easier because it's covered quite extensively. There's a lot of data sets that are available.

18:41What's interesting is historically, because of our history and background starting as a language learning app, We have spent a lot of time researching into that list of, you know, glossary term base that the professionals usually like to use within the legal setting, financial setting. And as a result of that, that has helped us to build, you know, that foundational source of getting that level of accuracy that legal teams require. The other part of it that's, I think Chinese is probably a bit easier because it's a language that's shorter in general. But when it gets to languages like Thai and Vietnamese, for example, when we were first dealing with it, the issue is how do you recognize the characters accurately?

19:33Because sometimes the placement of a dot within a character, especially for a scanned document, could be quite different. And then once you kind of like recognize it wrongly, then the entire sentence is contextualized wrongly. So you really need a very deep level of granularity to be able to recognize these characters effectively. And I think the same extends to an even more complicated language that we deal pretty well, which is Arabic. Because Arabic is written from right to left. And similar to Thai and Vietnamese, they have that character issue where you have, you know, like the placement of the dots, the combination of characters that needs a certain level of recognition.

20:22And that as a result, you know, we have spent a lot of time making sure that, okay, we recognize each of the characters correctly. Because for the longest time, I think what we have been hearing from quite a number of our clients based out of the UAE is that a lot of the issues of scan files is that none of these systems are able to detect the Arabic characters accurately. So when you translate it to English, it's kind of gibberish. And we have spent a lot of time fine-tuning and making sure our vision models are working really well so that we can deliver a high level of accuracy as a result. But that's one part of the observations, especially for character-based, non-Latin-based translations that we have seen.

21:11SlatorPod Host:What type of content would originate from Arabic-speaking countries, from UAE, going into Southeast Asia or Asia more broadly? Is it mostly on the legal side or would it be also consulting or investment banking material? It really depends. But at least based on our clients, a lot of it are agreements, especially, you know, they do a lot of cross-border work. And I think one interesting kind of anecdote that we heard previously was the issue of Arabic is that Arabic to English is already quite a challenge to find, like, good translators. And then if you're looking for Arabic to Chinese or any other language, the issue is what they will do is still translate it to English first and then from English to the other language, which, of course, during the process, you would see some disconnect, right?

22:03And that's where we can do direct, you know, like Arabic to Chinese or Arabic to Thai. And the time that it used to take was like sometimes a month for like, you know, a couple of 10-page documents just because you have to go through like different process and to make sure that it makes sense and to find the right linguist for such use case. and those are what we have seen. But also because of the increasing amount of trade between Asia and UAE and we start to see beyond that a lot of localizations of product fact sheets where they need to reach out to new markets to sell their products or services to.

22:43Those are some of the examples of where we see usage of our platform.

22:49SlatorPod Host:What are your thoughts on competition from what we at Slater call translation as a feature, basically embedded translation features in big platforms? For example, Adobe just launched, I think, Translate as a native feature and it still has the new thing above the feature. So it was relatively recently. But then, of course, you have the same with Google and Microsoft and everything. So is this something you're looking at very closely or you're feeling like, well, we're exceptionally focused on a specific set of problems and these big platforms will probably never get to the depth that we are? I think from a product building philosophy, what we always do is we are very much focused or obsessed over our customers and really solving the problems that they have.

23:41I think there are many of the big players who have been launching different forms of translation. Where we see why our clients have decided to go with us is really because we can tackle the entire suite of format types on our platform. So as an example, if Adobe, I haven't really looked that much into their product, but assuming that they would tackle most of the translation on Adobe type products and then there might be other, you know, document types that our client might need. And so that's one part of the solution offering that we offer. But the other part of it is also in terms of the understanding of the linguistic needs of the client and making sure that that's consistent across all the different file types.

24:35I think in today's day and age, even internally for us, when we are selecting products, we are trying not to get too many products at once. and I'm questioning if our clients, they're also not looking to get, okay, you know, like five different products for translation. They're just looking for that one that can solve across all their needs. And I do think that's where we come in to tackle that side of problem for them and do that really, really well.

25:02SlatorPod Host:Yeah, makes a lot of sense. So you started the company just kind of pre-ChatGPT, like most recent AI wave explosion. And then you said, you know, with the, you move from the language learning now to the translation part. Like, what impact did the kind of LLM explosion have as they became much more widely available to you? I mean, like, was this, or did you build the translation bit, like, after this was already extremely widely available? I think with the proliferation of the LLMs, that has also helped to overall increase the translation quality at scale and at speed for us. But from a market perspective, what has been really helpful is people are just much more open to the concept of adopting AI-related tools.

25:57In the past, I think, you know, people, you start to see the shift because you're telling them with something, they're like oh what about security you know is this AI can this AI be trusted but now people are just very hungry to adopt as much AI as possible because it has just become the standard for internal efficiency and people have you know really warmed up to the idea of AI adoption in general and that has really helped with adoption of our tool and to build the use case internally to management for them to say, okay, we have the buy-in and we trust that this tool is able to deliver at that level.