#261 Finding Product-Market Fit in Language AI with Naitiv Founder Gayatri Shahane

0:00When I actually started testing native a few months back, I would jump in on these Discord channels. They have these voice chat channels and I would jump in on those with Spanish speaking or Korean or Japanese. And a lot of the times people didn't understand that I was using AI.

0:23Gayatri Shahane:Hey everyone and welcome to another episode of SlaterPod. Very excited to welcome Gayatri Shahane on the podcast. She is the founder of Native, an AI agent for global expansion, and they're starting out with a live language interpreting feature. Hi, Gayatri, and thanks so much for joining today. Hi, Florian. So good to be here. Thank you. Always ask where people are recording this podcast from. So where are you recording this from today? What country, what city? I'm based in Davis, California. So it's one hour east of San Francisco. It's the home of University of Davis. So quite close to the source of this all.

1:00Gayatri Shahane:And yeah, we were just chatting before that you're also coming to the conference, to Slater Con Silicon Valley happening in about a month from now. So very excited to meet in person there. Good lineup. So, all right. So with Native, tell us a bit about the background before, kind of how the idea came about for Native and some of the other steps you had in your career before you started this venture? Sure. So my background is in a mix of business data science and AI. I studied at Carnegie Mellon and I worked in Bay Area startups as a data scientist in AI and business intelligence and have scaled teams and built data infrastructure.

1:43A little bit personal background, so I grew up in a family that did business across borders. So I was exposed early on to how much success in global expansion depends on people and culture. So I naturally gravitate towards solving these hard, meaningful problems that live at the intersection of people and technology and scale.

2:08Gayatri Shahane:At Carnegie Mellon, did you do anything in kind of language AI or was that something different? Oh, I studied energy science, technology and policy. And of course, data science, as an internal joke at CMU that no matter what you come in as, you come out as like a specialist in AI and data science. Because they have a super strong history also in kind of language AI and also machine translation very originally. So CMU is very well known for this. All right. So you kind of teased a little bit about where the impetus came for native, but tell us maybe the elevator pitch, the extended elevator pitch, and then a little bit more color to the idea and what motivated you to start this particular venture?

2:54So to start this particular venture has roots in my childhood. Like I mentioned, I grew up around my father's manufacturing business, where I saw he spent two decades actually expanding into 24 countries. I saw him navigate each language and culture and market and regulations and business norms. And after that, I also work in various startups who try to do the same thing, but then failing and trying things out and eventually working through a reseller who collect 40 to 60 percent of the revenue. So I know what an AI agent needs to let companies actually expand into other countries and to shortcut that 20 year process so that they can go in every country in days and not years.

3:45Gayatri Shahane:So when you talk to other founders, when you go from Davies maybe to San Francisco, Silicon Valley, what is some of the reception from the other founders when you speak about your idea? Like, what did they think about it? And the initial use case, as we saw, as we said very early on, is kind of that language interpreting feature. So what did they think about that? I think real founders always want to dig deeper and ask the hard questions. And they always ask them because they know what it takes. And then they want to see if you've done the hard parts. So they zero in quick with questions like how to handle turn taking or end of turn detection.

4:27What about the latest speech-to-speech research? What's your approach? And you know when you've built things yourself, you know what's supposed to be hard. So they're not essentially doubting you. They're checking if you've gone further than they can or they have. And when they stop asking questions and look satisfied, that's like their way of saying, all right, you've earned it.

4:54Gayatri Shahane:Did you have a co-founder or did somebody else help you start this? I have been a solo founder and I have a few advisors and friends helping me on this. allocated and dangerous. Yeah. Quite cool. So you speak eight languages is what we came across in our research. So how has this, like how it's been like almost a polyglot, very multilingual, how has this influenced how you think about AI and kind of AI translation in general? I grew up trilingual and I know three more languages and I tried learning two more, which is German and Chinese, in order to do business in those countries, but fail terribly because the languages are so hard for business communication.

5:43So I'm painfully aware that translation isn't communication. And you can be in a meeting for hours with an interpreter in it, but not quite understand if it went well or not. I was inspired by Erin Meyer's book, The Culture of Math, very early on. And in that, she talks about how cultures communicate differently and not just what they say, but how they make decisions, how directly they say it, and how trust is built in different cultures.

6:22Gayatri Shahane:I think that makes it super hard for AI generally, right? Especially in a live setting, because, you know, getting all these nuances right live is, I don't know, maybe it's even impossible. Exactly. So that's why when I think about AI and language, I think about all of these nuances and how to build this conversational AI agent that not just does work toward translation, but has that understanding of these deep cultural nuances. Because people don't just speak different languages, they do business differently. How would you do that? Like, how would you live feed that type of context into a system?

7:06Gayatri Shahane:Like, I want to talk maybe about maybe the technical side, but also the business model side, which is also hard, but let's go with the technical one first. So if you now assume, okay, there's all these complexities, these linguistic complexities, how would one live feed this into a system? It's pre-fed. So we do train the models and we have our prompt workflows, which are proprietary. And we train at per language specific pair, right? So all of these nuances are baked in into the language pairs that feeds the live model. And we also have different settings that the user can choose. If you're talking casually with a colleague, just choose casual mode and it will not add unnecessary formalities into that.

7:57And if you're talking to, let's say, 10 people who are Japanese and have that deep, respectful cultural background, you select the professional mode where your Japanese will come out as very nuanced and respectful with a very higher degree of honor fix.

8:15Gayatri Shahane:And when you say choose, like, choose where? Do you have a standalone app, a desktop app? Is it some, maybe like a Chrome add-on or how does it work technically? So the native conversational AI agent is a desktop app. And we decided that we needed a desktop app because we want to control a lot of aspects of it, which you can't do on a browser to control latency, to control the audio inputs, because a lot is going on in terms of audio input and output when you're doing simultaneous interpretation. We've just established that. I mean, obviously, it's super hard from kind of a problem point of view, a technical challenge, but it's also very hard to scale this as a business model.

9:02Gayatri Shahane:We've seen, we've been doing this for a while now at Slater and we've seen quite a few attempts at real-time AI interpreting. It's very, very hard to crack. We've seen many founders try and some actually give up. Why do you think you're going to succeed in this? Like what makes you confident from a business model perspective as well? Why I chose to work on this problem because it's still one of the few ones that's still deeply human and still unsolved. Like you mentioned, it's a hard problem to solve. And you can throw models at it and you can stitch together some APIs. And it will still fail because it's not just about the translation part.

9:45Sure, that's hard, but it's not just that. The main crux of it is the meaning. So when I actually started testing native, you know, a few months back, I would jump in on these Discord channels. They have these voice chat channels and I would jump in on those with Spanish speaking or Korean or Japanese. And a lot of the times people didn't understand that I was using AI. and we formed like real connections and made real friendships and people opened up to me and we were talking in Spanish and Japanese and Korean languages that I don't speak. So that's when I understood that we have something real and this will change the world.

10:36Gayatri Shahane:Okay, I'm gonna have to double click on that a little bit. So you're saying you, because I'm not familiar with that forum because I'm not hanging out in Discord. So how does that work? So you're on Discord, it's voice only, but you don't know. Yeah, tell me a bit more about that. And so you entered these and you were able to have a quote-unquote normal conversation using your own tech and communicating with a Korean, for example, but you were speaking English and that individual was speaking Korean. So was it fully live or was there a little bit of a delay or how did this work? On Discord, there are these language learning channels.

11:10So Spanish, Korean, Japanese, and thousands of people jump in and practice their Spanish or Korean or other languages. So it's not very, you know, restrictive. You're allowed to make mistakes and people are learning. So they have these voice chat groups in which you just jump in anonymously, you know, and start speaking with people. Sometimes there are two, three, four or more people in there speaking in that language. And this is live voice to voice. So at that point, they couldn't see my videos. I don't have to worry about lip syncing. And we did have real-time live conversations where connecting about stories and regions and learning about different cultures.

11:59It was surreal.

12:00Gayatri Shahane:That is so cool. So they were hearing an AI voice. And obviously, the setting was that it wasn't meant to be completely perfect. So maybe they were all thinking like, why is she in this group? She's really good. Right? Maybe there were some oddities, but probably it was pretty well done. Yeah, no, they were hearing my voice. I'd cloned my own voice. And it was coming out in a different language. And I would throw in some mm-hmms and ha's in the middle so that they wouldn't think that, okay. And laughs, because I think at that point, laughs are still not solved yet. but throw in laughs makes you like pretty human.

12:46And yeah, just jump on and start talking with them.

12:49Gayatri Shahane:Wow, this is incredible. Like if anybody had said that three years ago, it would have been, okay, that's crazy and now it's, yeah, that's where you kind of test your product, your initial product. All right, let's think about this as also a business eventually you want to make money with it and the competition is really, really strong, right? You have kind of big tech that wants to have this as a feature, Google integrating it and then at their kind of developer conference launching it. And Apple now also said they can do it on device, et cetera, but they never really can. And then you have like these large, what we call language solutions integrators, these big agencies that have billion dollar businesses and are kind of coming from the language services world where AI interpreting is an essential extension to their core business.

13:41Gayatri Shahane:So big tech and language solution integrators, lots of competition. So what are your thoughts generally on the competition in this space? Language is one of the oldest gatekeepers in business. If you don't speak the language, you're just out of the room. And big tech is solving it at the mass level. They're doing a great job of adding subtitles, auto-translate, auto-dubbing, post-processing. or even some type of like real-time wise translation. And it does help billions. But business is a little different. It's much more nuanced.

14:24And trust is the currency that we're working on in business and it needs a much more deeper understanding. And that's why we're not just translating. We're building an AI agent that's conversational and understand what it takes to do a business in other countries.

14:38Gayatri Shahane:Tell us a bit more about your early adopters, maybe like highlight the recent win if you have a kind of a B2B win or just some of the early adopters. Like who are these people? What are they using it for? We are currently working with U.S.-based B2B companies expanding into APAC, LATAM, EU, as well as Japanese bigger enterprises that want to expand into U.S.-speaking, English -speaking countries, or even Vietnamese Island and some other pockets of the world. So they're using Natives Conversational AI agent for live demos, onboarding, connecting within their team, global training, sales. These are typically sales teams, customer success teams, or partner channels.

15:25And anyone on the front line who deals in language trust building.

15:31Gayatri Shahane:I think Japan has historically been very, very interested in anything related to voice AI and AI interpreting. I've heard this from other similar startups as well. They're very keen on it if it works. And we recently had a success story, you asked me about that, with a Japanese company, like a US and Japanese company. native helped a us-based customer connect with their key account from japan which was on the verge of churn but that native helped it you know into expansion this specifically it was used in in-person meeting which was very critical and emotional because something had gone wrong And they wanted to convey those emotions with sincerity and, you know, respect.

16:24And Native did that beautifully. Usually the person who ends up being the bilingual trans, like the person who is bilingual ends up being a reluctant interpreter in that meeting. So that person told us that it would have been really awkward for him to deliver that message directly. And he didn't have to do it. and native did that and took care of that in a way.

16:49Gayatri Shahane:I love this. This is a new use case. Let the AI deliver the awkward or the bad news. Okay, that's a good one. It kind of puts a little bit of a distance between the two. Okay, so that requires obviously a bunch of components beyond the translation, like tone modulation, end of turn detection, speaker diarization, and all of this stuff. So how out of the box is this in like, you know, August 2025 for a founder like yourself? Or how much does that still need to be kind of built by you? So we have built our own voice orchestration engine from ground up because solving all these problems that you mentioned requires more than just good models.

17:34We use a lot of helper functions to optimize our entire audio pipeline. So for end-of-turn detection, we don't just like look at silence. We use a mix of semantic intent modeling. What language is it? Because pause detection is going to depend on which language is it. and and so a lot of helper models eventually trying to solve that that problem of end-of-turn detection because when you look at the the majority of the research on that right now is based for voice agents which is one after the other communication so end-of-turn becomes easier for that but when you look at interpretation you have to cut the other speaker jump in, do simultaneous, multiple speakers are speaking, and people are not used to having an AI interpreter.

18:29So they take a bit to get used to that, and they'll just jump in or keep talking. So we have to deal with all of these things and make sure that it's a smooth experience for the customer, that they get used to it very quickly, and that the native interpreter is not the key focus of the meeting and they just communicate with each other normally and the conversation flows. So that's the goal here. And for that, we build the voice orchestration engine from ground up.

19:02Gayatri Shahane:Did you see any difference in ease of adoption between different age groups? Like maybe, I don't know, younger folks having an easier time understanding, hey, here's an AI, let me kind of make sure that I let the AI follow or catch up? Or is it kind of the other way around? That's a very interesting question because our demographic, the key demographic is much on the older side because probably they don't want to learn a new language. And if they're going from like Japanese to English or other like Korean to English. And so we have to build for them and make sure it's super intuitive. So we are going towards that.

19:43But a younger person would just click a few things and figure everything out. And they're cognitive of things that I have to pause now. So those are the differences that we are seeing.

19:54Gayatri Shahane:You guys launched on Product Hunt a couple of months ago. So what are your thoughts on generally launching on Product Hunt? Does it help? Is it worth the effort? Would you do it again? I would say that Product Hunt is good for B2C and consumer products. so it's less relevant for B2B so if you're building anything for the consumer directly go ahead and launch on product hunt but for B2B it's not very relevant we did it for testing, messaging and doing a launch felt really good that's a good point also one thing that you're maybe facing already but that you will definitely face in the future is building a product, having the vision and then catering to like a bunch of enterprise customization requests once you start getting the big leads and once you have like your initial key customer who wants you to do all kinds of custom stuff.

20:52Gayatri Shahane:Like what are your thoughts on that? Like sticking to the vision versus, you know, occasionally catering to customization or being like fully open to any customization? Oh, it's always a balancing act, right? and we have had to focus on security and compliance to even get a foot in the door with these enterprise customers. That's the first thing they see even before the product. So it is important if you're going for B2B enterprise sales to have those things ready. In terms of customization and requests, we get a lot of great feedback from them. I mean, I was just at the day, I was on a call with a Japanese company and they had 20 line items with feedbacks and what they want to do, have in the product, which is great.

21:44I love that. If they're asking us to build something that makes the platform better for everyone, then yes, great. It goes on the product roadmap. But if it's like too bespoke or too abstract, we put it on a separate track or, you know.

21:59Gayatri Shahane:All right. So you're here talking about the product. So this is, I guess, part of a little bit also helps with marketing, but like your sales and marketing, what are your thoughts there? Product led mostly, like you're building the most amazing product and then people will come. Do you want to maybe try to engineer some type of viral moment at some point or maybe do a little bit of extra marketing? What are your thoughts to growing the product? We haven't done any ad spend or marketing yet. It's all founder led right now and sort of organic. One great thing specifically about the conversational agent is that when a company uses it in a meeting and the experience is great, the other participants who are there, they're like, oh my God, what is this?

22:45I want it. That's how I've gotten my current customers and built the pipeline.

22:51Gayatri Shahane:So I guess that would qualify as product-led growth. Like the product speaks for itself and then boom, boom, boom, it goes on. So do people, when they use the product, is there some kind of branding somewhere? Would the other participant understand that they're on this because they have to download the app or how does that work for the other, not the user but the person that's getting invited to the call? So we do have a couple of modes that you can use the conversational agent. we used to support online meetings as well as in-person meetings. For online meetings, the first mode is the native bot mode, which you would just invite it as a participant on your call and do consecutive interpretation.

23:44This helps when there is a mix of people, you know, there's like 40 % English speaking and 60 % other language, and everyone needs to know what's going on. That's great for them. But we also have the desktop app, which is simultaneous interpretation, in which I'll just keep speaking and you keep hearing me in your language. And you probably won't even hear me in my original language.