Full Transcript

Dhruv Shama: Hey there listeners. How is Wednesday coming along for you? We're streaming live. This is stream 107. Today Utsav and I are chatting with Dhimant who is the founder of a voice AI company called Maya Research. Dimant, welcome to The Offline Network. Hi, how are you?

Dheemanth Reddy - Co-founder & CEO, Maya Research: Thanks for having me. Hi, what's up? Hey, Dimant.

Dhruv Shama: All right. Dimant, let's just get started by having you describe Maya Research in your own words.

Dheemanth Reddy - Co-founder & CEO, Maya Research: Yeah, at Maya Research, we're building a voice interface where the next 5 billion people will interact with phones by just talking and can access anything from payments, shopping to utility by having conversations with AI and the things that are happening on the screen.

Dhruv Shama: Interesting. And you have the app and then you've also trained an open source model.

Dheemanth Reddy - Co-founder & CEO, Maya Research: So, we started our journey with an open source model to show the world that what we can build and that India can also build open source voice models that can beat Frontier. So, and basically the interface, the app is a voice model, right? At the end of the day, voice interface is you're talking to your voice model. So, you can boil down the company to a voice model. That's what we do. We build conversational voice models that can speak like a native person. And the first model that we released is on par with the GPT's real-time tool model in the speech- learning leaderboard. So, and also Maya is the only company from India to be on the leaderboard.

Utsav Somani: And that's fascinating because you've only raised, I think, from South Park Commons, 1.9 million. But I'm guessing you might have to answer this question a few times in your conversations with investors and employees even. Like, what is your true advantage? Because you've raised so little. There are many other voice models out there. So, what truly is your moat and how are you defining it right now?

Dheemanth Reddy - Co-founder & CEO, Maya Research: I think three things. One is, I think building voice is something instinctive to us. Me and my co- founder grew up in small towns of Andhra Pradesh. We kind of grew up in places where voice itself is the mode of communication. For us, text is new. For us, voice is the thing that's original to me. So, when training the models, those things come up in the architecture. Like, how does the rhythm should sound? What's the timbre? What type of data we should bring in? I think that level of understanding helps us to make quicker and faster decisions so that we don't burn a lot of money on experiments. A lot of big companies spend a lot of money on experiments, right? They spend a lot of money on whether this architecture should work, whether that architecture should work. We don't do that. And second is, we have a data player already, right? Our app generates so much data for us that we don't have to buy it from other people. We have a data advantage that our app is live in entire 226 towns and we get so much data to train our models. And third thing is, I think I intentionally put the company in low resource mode because that's when you innovate. Getting early on capital of, let's say, $30 million or $40 million, your mind quickly goes to, okay, where should I buy the data from? Which lab? Having less money helps you to innovate faster. I think these three things, like, you know, combinedly help us to do frontier models at a very low cost.

Utsav Somani: And talk to us about Maya 1, which is your 3 billion parameter model as well, and the open source and why these decisions?

Dheemanth Reddy - Co-founder & CEO, Maya Research: One, I think I was frustrated by the open source quality of models outside because everybody had a certain number of speakers. It was so frustrating to me because the voice is such a personal thing, right? Like, imagine a world where people are talking to eight voices. That's a disaster. Like, voice itself is very personalized to us. So we built a model that can generate new voices. So Maya 1 is especially good at creating a new voice. You can go and type, okay, I want Australian accent, but should speak Hindi really well. And American accent, but should speak Australian words really well. So I think that's what we did. And the reason why open source, to be honest, it's really fun. Like, we just wanted to show what we could do from India, and where this will go. And the amount of downloads. So basically, Maya 1 got 440,000 downloads in the last six months. We haven't expected that. The biggest driver for that is ability to create new voices.

Dhruv Shama: Can you share some stories of applications built on top of your open source model?

Dheemanth Reddy - Co-founder & CEO, Maya Research: Majority of it is, again, customer care. That's why we started building our app very early on, because we haven't seen early adoption of what people are doing with voice models. Majority of it is customer support. Second is content creation. Third is companion. And very less into utility. And that's what we kind of observed very early on, because till December at Maya, we were doing open source conversational models. At December, we thought, see, voice is such a generational shift of how humans will interact with technology itself. People are not going to type from now. People will interact with robots, phones, laptops, everything. So that helped us to build what could be something that people will love to build on top of conversational models. Something very utility focused.

Dhruv Shama: Deepak, when we go from a typing-first world to now a voice-first world, and that, by the way, is a settled argument. There's no denying it. But I would say that all of us would pretty much type in the same way. I don't think there was much of a difference between how a Chinese person typed and how an Indian person typed. But we all speak in different ways. So what's so unique about Indian, like what's so hard about cracking voice AI for India?

Dheemanth Reddy - Co-founder & CEO, Maya Research: I think it's context. More than the rhythm and timbre, Indian languages have a lot of context delivered through voice. In the West and in different languages, people who are typing first, they have very less context delivered through voice. When they say something, they mean the same thing, actually. In India, when they say something, it's actually completely different of what they're trying to mean. If a conversation between a shopkeeper and a woman in a tier-three city, if she is silent, that means it's not a no and it's not a yes. You don't know, really. You need to look at her expressions. But probably very experienced shopkeepers can tell what she's saying. I think a lot of context is missing. It's because there is no data that's in the Internet or with any other lab that what does this context mean? When they are talking, I think that's the key to crack. That's what we're doing at Maya, to understand what people are saying and what do they mean by that.

Dhruv Shama: You're saying when Maya 1, if someone asks me if I want to have a gulab jamun, I say, no, no, no.

Utsav Somani: All the Indian husbands who are tuning in right now, they'll be like, shit, man, I need this for home as well. Whenever my wife says something, she probably means something else. We need more context. Three million downloads.

Dheemanth Reddy - Co-founder & CEO, Maya Research: Telugu being your biggest base right now? Uttar Pradesh and Telugu are our biggest bases right now. Telugu is because we completely understand the GTM part too. In the north, Uttar Pradesh is our biggest market. In the south, it's Telugu.

Utsav Somani: You've mentioned previously that your enemy isn't the competition, but I think latency is. Elaborate more on that.

Dheemanth Reddy - Co-founder & CEO, Maya Research: See, voice is such a very interesting part that you don't want to wait for a lot of time. Think of it this way. You're talking to your girl, and if she's literally taking so much time to talk with you, you feel that hesitance. You don't want to stay there probably. It's the same thing that goes with everyone. Latency kills conversations. All voice AI companies are fighting with latency. When they talk to AI, AI should acknowledge. AI should respond faster. That keeps you to be in longer conversations. Right now, in our app too, people come to Maya and have conversations. Some will figure out having lengthy conversations. People who can figure out low latency conversations. Some people churn off faster because it's not natural for them. They live in a world where people are continuously talking. Indians interrupt a lot. That's crazy. But if voice AI is taking 2-3 seconds to even talk, that just kills the experience of not going more than 10-20 minutes.

Utsav Somani: Architecturally or technically, what decisions are you making to solve this in your own way?

Dheemanth Reddy - Co-founder & CEO, Maya Research: It boils down to when to talk and what to talk. The latency part is always about not just throwing off some random words. You need to know when to talk. Based on the conversations, we are training the model to figure out what is the right time for you to enter the conversation. Like in 200 milliseconds or 400 milliseconds. And what to talk there. Should I say, hmm? Or should I say, uh-huh? Those two, when and what, is what we are training for right now.

Dhruv Shama: Most of us Indians are also multilingual. Or at least we use English plus one other regional language. We'll keep switching between the two. I want to understand from you, what is code switching? Why is the technical price for solving that very high? Talk to us about that.

Dheemanth Reddy - Co-founder & CEO, Maya Research: I have a very contrarian view about this. I don't think code switching is a problem at all. I saw a lot of voice AI companies literally posting that, okay, we can do code switching. But code switching is not a problem. What's the biggest problem is whether the voice that you are speaking to, that you are using is native or not. The problem is, if you try to solve code mixing, I'll tell you how. If you learn Telugu, you will never sound probably like a native guy. Even if you learn code switching in Telugu, like if you're speaking Hindi and you add a bit of Telugu words, basically that's what everyone is calling code switching.

Dhruv Shama: If you're throwing me a challenge, then give me 10 years, I'll return with a native Telugu accent.

Dheemanth Reddy - Co-founder & CEO, Maya Research: Perfect. Okay, so that code switching is what people are talking about. But the problem here is, all the big AI labs are using the same voice. They're using the same voice to speak 11 languages. That's a bizarre combination. It's super, super hard to solve for that because you need one voice to speak like a native person. So that's why at Maya, we take Telugu voice and make it only speak Telugu. Probably some code mixing of English. Because if Telugu voice starts to speak Hindi, it doesn't make sense. It doesn't be that way. And code mixing, code switching is generally have like one or two languages. For South India, it's always about English words and a bit of Hindi. South Indians barely use any Hindi word switching. In North India, it's always English, right? You guys don't use South Indian languages. So it's a very small problem. And I don't know why people are talking so much about that part.

Dhruv Shama: So you're saying if it's a TTS model, then it's a problem. But if it's a speech native model trained on Indic language data, I'm just using that term generally, then it's not even a big...

Dheemanth Reddy - Co-founder & CEO, Maya Research: Yeah, it's not a problem at all.

Dhruv Shama: Not a problem. Okay, cool.

Utsav Somani: And have you started thinking about a business model? Enterprise clients are mostly focused on the app and improving the model right now.

Dheemanth Reddy - Co-founder & CEO, Maya Research: It all boils down to the model. So right now the current focus is on consumer, like scaling the app. And of course, we have the enterprise motion branding. So MyEyes models are available in Phal, the biggest inference provider in the West. So people can access through that. They will be live in AWS, Azure, and some voice agent companies too in the next three months. So people can directly plug into their enterprise workflows.

Utsav Somani: And what are some of the challenges that the lab and the team are looking forward to solving in the next one year?

Dheemanth Reddy - Co-founder & CEO, Maya Research: The next one year would be... We have a bar set in the company that... We are driving the models towards the most human conversations. I'm completely obsessed with... The voice AI will take off completely differently when it can sound and speak like a real human, just like how you and me are talking. The amount of conversations, the level of conversational ability will decide whether voice AI will be the future or not. So it's always about the humanness of the model. That's what we are scaling towards. In the next one year, you'll probably see Maya 3 or Maya 4, which can really, really have low latency Hindi conversations that you basically feel like you're talking to a very, very real human.

Dhruv Shama: Even though you started by saying you're not building Maya for the 50 million top-tier consumers in India, you're building it for the 5 billion people. But stretch our imagination a little bit about how they're actually going to interact with voice and make it a part of their everyday lives.

Dheemanth Reddy - Co-founder & CEO, Maya Research: I think the reason why we are so obsessed with the next 5 billion is we have observed these 5 billion people growing up that they haven't experienced any technology. For them, shopping is still a big thing. For them, downloading a document from the internet is still a big thing. So for the first 50 million users, people who are ordering in Zepto or Zomato, they know. They know how to do things. They're using MCP, they're using Cloud. The next 5 billion people, that's a complete way of interacting technology altogether. Imagine if you put that technology in your grandmother's hand and she is talking and she's doing all the things that you are doing. That's a great future to be in. And I think I'm so scared of not living in that future. So that's why we're driving the company towards that.

Utsav Somani: All right, Deemant. Thank you so much for coming on to you and wishing you all the very best. Thank you, folks. All right, folks, we're moving on to our next guest. We've got Mridu from Terrarium with us. Mridu, welcome to the show.

Mridu Jhangiani - Founder, Terrarium: Hi, Utsav. Hi Dhruv. Thank you for having me.

Utsav Somani: Can we introduce what the company does?

Mridu Jhangiani - Founder, Terrarium: Yeah, sure. In a single sentence, Terrarium is a platform that accelerates the adoption of sustainable materials by providing founders with access to knowledge and networks.

Utsav Somani: And I mean, for our listeners, for the benefit of our listeners, what are sustainable materials and alternative materials? Describe them a little bit for us.

Mridu Jhangiani - Founder, Terrarium: Sure. So I think one of the challenges actually is that different people use different terminologies. But if I had to break it down into three categories of what usually confines the definition of a sustainable material, the first thing is the choice of feedstock, right? Is this coming from plants? Is this agricultural waste? Is it circular, essentially? The second category that I would say is performance. So when we use the term next-gen materials, we're thinking about, okay, is this better than the incumbent material? What additional features and properties does it have? And the third category, I would say, is the net lifecycle emissions, where we're trying to understand, okay, manufacturing to end of life, has it reduced the carbon footprint by a certain amount?

Dhruv Shama: Mridu, maybe I'll ask you a very fundamental question. Why do producers go looking for alternate materials in the first place?

Mridu Jhangiani - Founder, Terrarium: I mean, I think there are two primary drivers. One is, of course, regulation. You know, especially in Europe, you're seeing a lot of regulatory push to be able to, at least for the large companies, to report publicly on how they're reducing emissions. And one of the key ways that they do it, especially for sectors like fashion, for example, is the only way they can actually achieve their net zero commitments is by reducing the material consumption or by switching to an alternative, more sustainable material. That's one thing. And then the other thing with corporates, I would say is that at the end of the day, the voice of the customer, the demand of the customers are sort of driving force for many of the decisions that they make. So when consumers are becoming more aware of sustainable alternatives, it's also pushing corporates to think about, okay, how can we actually provide for it?

Dhruv Shama: Mm-hmm. And when someone innovates in the realm of materials, why does it very often take that innovation, you know, lots and lots of time to see the commercialization? Like, why is that pipeline so broken? Why does it take so much time?

Mridu Jhangiani - Founder, Terrarium: I mean, in many ways, the, you know, the similarities that I've seen is with any deep tech startup, right? You need a certain amount of patient capital. You need access to a certain kind of talent pool. You need to be able to sell your vision. And a lot of times, you know, founders are great when it comes to science and technology, but to be able to sell your vision, to be able to communicate how your product is better, why it's better, can be a difficult thing. And sometimes brands focus not to choose on it, choose to do that. So that's another reason. There's also a lack of, I would say, standardization when it comes to the kind of certificates that you need for compliance. And think about it, if there's a completely novel material, sometimes it doesn't even fit that category or bucket or checklist that exists for, you know, the most prevalent materials. So I think there are lots of different challenges, and I think it'll take the ecosystem to come together for a lot of these materials.

Utsav Somani: Can you give us some examples of how the ecosystem has come together, or maybe some projects that have seen the light of day? Any sustainable projects that you think were exciting in the alternative materials space?

Mridu Jhangiani - Founder, Terrarium: I think one sort of difference that I'm seeing in how startups approach this whole scale-up process is earlier there was this sort of affinity to only choose large brands so that you could have large brands and show traction and credibility by engaging and having a pilot with them. But one interesting pivot that I've seen is now that material startups are also looking at building that traction and credibility with smaller brands, smaller projects. It could be one-off demonstration projects. It could be with smaller, more niche brands, just different ways in which you can build that credibility so that when you approach a big brand, a large corporate, you have something to show as proof.

Utsav Somani: You're losing your audio a little bit, Mithu. Yes, I think so.

Mridu Jhangiani - Founder, Terrarium: Okay, I hope I didn't have a funny case.

Utsav Somani: No, no, you have to repeat the last, I think maybe a minute or so.

Dhruv Shama: What happens with the voice packets is though they get compressed and then they deliver all at once. So we heard a lot of what you said, but in three or four seconds, you don't have to repeat yourself.

Mridu Jhangiani - Founder, Terrarium: Yeah, just to recap, we're basically not going after the big brands and the big brand names, but looking at smaller projects. There could be one-off demonstration projects. There could be with a high-end fashion brand where you can have your proof of concept basically on the runway.

Utsav Somani: Can you give a name of these companies and founders who are doing exciting stuff?

Mridu Jhangiani - Founder, Terrarium: Yeah, I mean, I would give a shout-out to two of my favorite founders in this space. One is Amrita. She's building a company called Amilo. They use basically plant-derived starch to create biodegradable and compostable packaging. And the other one is Sangeeta. She is the founder of Hedrad Panels and they're using recycled paper to create panels that are non-load-bearing.

Dhruv Shama: That's very cool. Mridu, you know, we once had a founder on this show. His name is Sripur and he's building a company called Arcturus Aerospace. And so, you know, autonomous drones. And he told us how important the materials are to what he's trying to do, right? Because, you know, like stuff has to fly. You want it to be light. You want it to be strong and so on and so forth. He's building an aviation company. He's not necessarily building a materials lab. Are there, you know, spaces and experts who founders can be speaking with if they want something done but they don't want to become materials lab process?

Mridu Jhangiani - Founder, Terrarium: Yeah, I think there definitely are. A lot of them you'll find in the industry with years of experience that have seen manufacturing scale up, that have seen material choices change. But I think the problem is access to those industry experts, right? I think there's an interesting, like, shift where we're seeing senior executives' leadership at some of the, you know, metas of the world that are now looking to come onto advisory boards of some of these startups. And then if they feel like it's an interesting, you know, alternative or a safe bet, then move into these startups full-time. So I think it's definitely possible. But again, since we're talking about materials, we're talking about materials that perhaps may have never existed. I don't think there's a defined, like, pathway or roadmap to scale or any certainty as such that can be guaranteed. It's just being in that discovery phase together.

Utsav Somani: And you've done something with climate designers as well. I think the first Asia chapter for them. What is that initiative?

Mridu Jhangiani - Founder, Terrarium: This was years ago, actually, during COVID. And I landed to, I mean, I had a lot of time on my hands. And, you know, sustainability has been an area that's close to my heart for as long as I can remember. And it was just me trying to bring together a group of people, give them, you know, a social outlet during COVID. And the kind of topics we covered was like home composting and, you know, avoiding recyclable, non-recyclable plastic. So, yeah, that was actually in the distant past.

Utsav Somani: Talking about one of the most polluting industries in the world, is fashion still number one? Or are there other heavy metal industries which are taking the lead of what?

Mridu Jhangiani - Founder, Terrarium: Yeah, I mean, definitely the construction space is a huge one. Steel, cement, that's a big one. But fashion also is huge. Plastics as well.

Utsav Somani: And what innovations do you think in the fashion space will sort of help reduce this going forward?

Mridu Jhangiani - Founder, Terrarium: Would help produce, like reduce the waste?

Mridu Jhangiani - Founder, Terrarium: Yeah, waste innovations, yeah.

Mridu Jhangiani - Founder, Terrarium: Yeah, I think on the material side, it falls into two categories. One is the actual product replacements. So you're seeing startups that are working with banana fiber, hemp, pineapple waste to develop new direct product replacements. But then you're also seeing some that are innovating at the technology level. So, yeah. But I think the majority of startups that I've seen are basically using some kind of natural material and converting it into a direct new fiber that can be sent as a drop-in solution to many of these fashion brands and their manufacturers.

Dhruv Shama: You're a designer by education and by trade. And we had a designer, you know, someone that designed for an academy on Monday as well. So we're going around asking a lot of designers this question. How do you see, you know, your trade, your craft changing in this era of AI? What are you guys excited about? What do you guys worry about?

Mridu Jhangiani - Founder, Terrarium: Yeah, I think one exciting thing, especially, so my specialization is in digital product design. My bachelor's was in product design. And one exciting shift is how AI actually helps you collaborate with PMs and engineers in a way that we haven't seen before. And so on, you know, some people see it as a con that the lines are getting blurred. Everyone can have everyone's job. But I think the pace at which you're able to iterate, move, build in features, it actually gives you a lot more time to think versus, you know, be within the frame, pixel pushing. So, yeah, that's what I'm excited about.

Dhruv Shama: You no longer just hand over a Figma file, like our guest on Monday told us, you push directly to the client's GitHub. So that's been one big change. The lines are blurring.

Mridu Jhangiani - Founder, Terrarium: Yeah.

Dhruv Shama: Yeah.

Utsav Somani: And talk to us about your time at Meraki Labs, which is a venture studio, right?

Mridu Jhangiani - Founder, Terrarium: Yeah. Yeah. So I joined Meraki Labs as a digital product designer. But what drove me to join Meraki was, you know, as a design student, they often don't teach you a lot about business. I remember in my classroom, we had a lot of great projects, thesis projects, but, you know, that was it. It kind of was left behind in the classroom. And there was this question at the back of my head, which was how does an idea become a company? You know, what else does it take? And so that was actually my trigger to move into this world of venture studios, startups, and understand and learn whatever I could. Journey was exciting. I learned a lot, had a great team to work with. And in fact, the skills and the sort of insights that I picked up from there is what led to Terrarium.

Dhruv Shama: Fantastic. Mridu, maybe as a final closing question, we started with materials. Maybe we can end with materials as well. What are some materials worth paying attention to in India and abroad? What materials that founders very, very early in their design journey can take advantage of?

Mridu Jhangiani - Founder, Terrarium: Personally, I'm quite excited about the packaging space. I think there are so many different startups that have taken a different approach, a different material input to create alternatives to films, to, you know, the lining that you have inside burger boxes, to create the replacement for single-use plastics. I'm most excited about that, one, because I love being on the beach and when I see a lot of plastic, it really gets to me. So that's an innovation where I see there's a lot of variety in the terms of solutions that are coming up.

Dhruv Shama: Yeah, I know this one hits home because when you drive through our cities, I think the number one, Amritsar was asking that question about who emits the most amount of pollution. If you ask, like, if you're looking at individuals, people, it's mostly just packaging boxes. But yes, what arrives in Zomato and Swiggy deliveries is starting to look very, very different from three, four years ago.

Mridu Jhangiani - Founder, Terrarium: Yeah, absolutely.

Utsav Somani: All right. I think that's a good note to end the show on. Thank you so much for coming on our show. Thank you for joining us.

Mridu Jhangiani - Founder, Terrarium: Thank you so much for having me.

Utsav Somani: All right, listeners, that's it from us. We'll see you on Friday at four o'clock, same time. Bye-bye.