Full Transcript

Dhruv Sharma: All right, listeners, welcome to our first live stream of March. We've got four guests today, each more interesting than the other. For our first guest, we've got lovely Amrita who's joining us from Darjeeling. Apparently, she's at a company off-site and she'll pop on the screen in just a bit. But she's written this fantastic piece that we'll talk about that, in a sense, marries the most modern technology of the time with one of India's best-known and historical textiles. And with that, let's welcome Amrita.

Utsav Somani: Amrita, welcome to the show. We had Shrey from Alt-Carbon as well, and he spoke to us about Carbon Crest, what Alt-Carbon does. But you're behind Alt-Mag, and this piece that really stood out, I mean, of course, the fascinating arguments that you made, but also, I think, the design element of this whole thing. So we wanted to spend the next five, seven minutes with you talking about the core crux of this article. The article is Kanchipuram Sarees and Thinking Machines. And there are, I think, a three-part existential crisis that you highlight for the silk weaving industry. So do you want to describe that and start us off with that?

Amritha Sreekumar — Alt Carob / Alter Magazine: All right. So the narrative of the article talks about some of the social issues and the challenges faced by both the artisans and also the industry in general throughout the years. And it talks about technology as a tool. It doesn't post it as a main solution, but it talks about how different activities can address different problems facing the industry from, say, traceability, like blockchain to identify and champion the artisans to the end users, to microbial dyes, which basically improve the waste generated by the industry and also the health of the artisans, from synthetic dyes, which can be carcinogens, and finally, to the use of AI in improving or speeding up the design process itself, which is suffering a crisis both due to the fact that younger people are not joining the industry. Like, if you look at the average age of a Kanchipuram weaver, it's 38 years old. And it takes a lot of time to upscale somebody to the point that they can generate, not in the sense of AI, but by themselves create designs that are not just traditional, but also flexible enough to adapt to the modern. So we've been looking at AI as a tool that can fill some of these gaps while the industry goes through a process of social improvement, like through governance, be it through government intervention or through other investors or training, but we are posing AI and technology as a support structure.

Utsav Somani: And do you think, I mean, you mentioned in your article as well, that young people are not joining this trade. Will the use of technology change that?

Amritha Sreekumar — Alt Carob / Alter Magazine: See, there are two parts to this answer. Technology is not a be-all and end-all solution, but at the same time, these are very traditional industries. They move slowly. If the information or if this traditional knowledge dies out, to revive it from an absence is very difficult. Technology is posed as a stopgap solution here. It's like some of the challenges that weavers face that they cannot solve immediately at this moment in time, can AI take some of the brunt of that? Can it learn the grammar of this design so that it is preserved in a way that is not just human memory? So in the meantime, we are not saying that other interventions or other solutions shouldn't happen. This is a support or it's a support net, basically.

Dhruv Sharma: How old is the Kanjivaram Sari Amrita? Do we know when the first ones may have been born?

Amritha Sreekumar — Alt Carob / Alter Magazine: Actually, basically the Adai weaving loom, it's a millennia old. We don't even know when the original handloom came about for sure, but it's survived a lot of things. So I don't think it's an industry that's going to die out in the modern age.

Dhruv Sharma: And the use of AI at this point in time, is it limited to the design process?

Amritha Sreekumar — Alt Carob / Alter Magazine: So if you look, if you check out the AI Summit, GCS actually came up with an AI which uses light technology and artificial intelligence to support the weaver. So it detects errors. So as you weave, right, if there is an error, you miss a particular weave, it will highlight it then because like a lot of time is wasted if you miss, you complete the design with the error in place and you have to remake it. So that is one way that AI can be used.

Dhruv Sharma: It's kind of like the typewriter. If you make one spelling error, you have to start from all the options.

Amritha Sreekumar — Alt Carob / Alter Magazine: Exactly. So, but the way we were examining it in the article was a little different. For instance, suppose you want to come up with an idea for a bird based, if you look at Kanjivaram motives, it's mostly based on nature and temporary designs. So there's a lot of spiritual motives as well as nature motives. Suppose I want to explore with say the nightingale. It is not a traditional bird that you would see in Kanjivaram. So instead of like a team of five designers working on options of nightingale translated into geometric patterns that can be woven for like two weeks, you can speed up this process of exploration using generative AI.

Dhruv Sharma: You're saying that saree has always been like a canvas, a medium for storytelling. Every symbol has a meaning.

Amritha Sreekumar — Alt Carob / Alter Magazine: Every symbol has layered meanings. For instance, if you look at a royal procession of horsemen, elephants, etc. From one angle, it describes prosperity. From another angle, it describes authority. That kind of a saree won't be worn to a fertility ceremony, for instance. That would be like after marriage or during pregnancy, the kind of saree that would be worn to celebrate that would have manga motifs, which is a symbol of fertility. So each of it has cultural context and deep layered meanings.

Utsav Somani: And how do we get a corporate fortress, like say Prada was talking about the Kolhapuri chappals in their fashion line, right? How do we get something like enormous or like a luxury house built out around these Indian artisanal traits or skills that we have as a country? What would it take?

Amritha Sreekumar — Alt Carob / Alter Magazine: Sorry, it's a question whether to champion our craft through such brands or...

Utsav Somani: Or even build a brand around it or even build a brand around it.

Amritha Sreekumar — Alt Carob / Alter Magazine: I think, see, the thing is that because the craft itself is valuable, it is rare, it is an art form that you're wearing. Brands will always find it appealing, especially foreign brands. So I think while we discuss this a lot, the perspective we took in the article is how can we champion the artisan themselves to the end consumer without any of the retail houses or middlemen taking the credit? Like suppose I buy a Kolhapuri chappal from Prada, I don't know the name of the artisan who made it. So a lot of people say when we suggest blockchain as a solution, they are overengineering the problem. GI already does the task of ensuring authenticity and traceability of the designs. But what does blockchain potentially do? It can actually take the name of the artisan, like a mangakaar. You know their Twitter ID, you know their name, you know who is telling these stories. Why should stories told on saris be any different? I want to know the name of the weaver who wore the sari I'm wearing for my wedding. It's like a once in... I mean, most people want it to be a once-in-a-lifetime thing, right? So that is what we are trying to achieve through blockchain. And there are other brands, maybe not in the textile space, like if anybody checks out Tony's Chocolate Only, it's a British-based chocolate brand. They ensure traceability in the entire logistic process from sourcing the cocoa to the end product. They ensure that there is no human exploitation, you know, the farmers' children go to schools, like education, that they are not forced to come back.

Dhruv Sharma: Fair trade practices.

Amritha Sreekumar — Alt Carob / Alter Magazine: Exactly, fair trade practices are ensured through blockchain traceability. So this is something that you are suggesting is possible.

Dhruv Sharma: Anytime you're buying an artisanal product, which I mean, they really matter now, especially in the age era of AI slop. If you're really holding with your hands something that's a truly artisanal product, it could be, I don't know, it could be bread, it could be a sari, it could be a Cuban cigar. You want to be able to trace its provenance and blockchain makes that very easy to do.

Amritha Sreekumar — Alt Carob / Alter Magazine: Exactly, exactly. And it doesn't happen easily, you know, like the thing is that when we talk about this, we have to talk about the cost of building the infrastructure. It can't be done, it can't be championed by the handloom industry and their unions and the artisanal groups themselves. You need the government to step in at some point to build this infrastructure. So that is a challenge that these solutions pose. It can't always be like, I don't like that it's an external solution, like we want a tech company to come build this for artisans. Like it would be really nice that we could create a system where it is co-created with artisans themselves.

Utsav Somani: Awesome, Amrita. Thank you so much for coming on the show and educating us. Thank you. Thank you so much. What can we look forward to from the Automag next?

Amritha Sreekumar — Alt Carob / Alter Magazine: We are focusing on telling deeply researched long-form stories, like we have a few in the works. I think our next one is about to be about, it's going to be fun, it's going to be about ice cream. So please look forward to the stories.

Utsav Somani: You've timed it for the summer. I'm in summer season, yeah. All right, enjoy your offsite in Darjeeling. Thank you so much.

Amritha Sreekumar — Alt Carob / Alter Magazine: Thank you. Thank you for having me.

Utsav Somani: Welcome to the show. Thank you guys for having me. Awesome. So let's introduce Guarantee to our listeners. What do you do?

Tanay Goel - Garantie: We are a product company with strong tech capabilities and brilliant distribution spread across the country. So we kind of help OEMs with value-added services, which is in automobile and mobile industry, products like extended warranty, AMCs, protection plans for your mobile phones, etc. So that's the product. Then we build tech to integrate the entire sales distribution and service network. So that's the tech. And then we have great distribution arm or spread across the country to enable sale and sales for that. So if I have to put it in 30 seconds, this is what we do.

Utsav Somani: And what's the scale of the business right now?

Tanay Goel - Garantie: We transact or acquire around 400,000 customers a month. So I think we'll end up this year acquiring 5 million customers. Yeah, so that.

Dhruv Sharma: And what's your definition of a customer? Is it distributors, retailers, the whole chain?

Tanay Goel - Garantie: No, no, no. It's the end customer. So it's the customer who's actually bought, for example, that two-wheeler and has bought our product. Or it's the customer who's bought that mobile phone and has bought our product. So that's the end customer for me.

Utsav Somani: And extended warranty is not a problem that most people think about solving. It's an unsexy business. How did you discover the problem or why was it something that you were uniquely positioned to solve?

Tanay Goel - Garantie: Yeah, so you had the answer in your question. It's I like unsexy business. So we entered this space six years back, especially in the automobile space. And before that, I used to run an insurance-broking company, which I sold. And we were thinking, OK, what next? What next? What problem should we solve? And that time, Policy Bazaar, CoverFox became very big. So I was trying to pivot and we saw an opportunity in this side. And we saw that, you know, the audience were busy reaching out and setting the distribution business in India. So their only focus was market share. And that's where we entered there. OK, you guys keep focusing there and we'll build value around your customers.

Dhruv Sharma: Maybe I'll quickly share a story and you can share similar ones straight from the field, right? I remember once someone told me that the automobile aftermarket OEMs that sell, I don't know, headlights and brake shoes and so on and so forth, sometimes on their box will have like one of those scratch codes so that when it reaches the garage, the mechanic can scratch it and avail some sort of an incentive. And that gives them an added incentive of recommending one commodity product over the others. Do you see other examples like that in the industry as well?

Tanay Goel - Garantie: See, my focus has been really to add value to the OEM. By that I mean, if I add, how can I uplift their brand? How can I uplift their reach to the customer? How can I make sure that the customer stays in their ecosystem? How can I make sure that the customer goes from their one bike to the other bike? How can I make sure that he never leaves their ecosystem? That's my job. So I look at things from that lens. So there's nothing short term that we do. If we introduce a product, we're looking at a life cycle of that customer with the OEM for next 10 years.

Utsav Somani: But OEMs could pretty much build all of this in-house, right? I mean, at least the extended warranties and some of the OEMs do offer that.

Tanay Goel - Garantie: Yeah, so extended warranty is a very basic product.

Utsav Somani: So the revenue split for you is extended warranty one of the biggest?

Tanay Goel - Garantie: In the two wheeler space, extended warranty became very big. Then we actually, we had integrated their entire service network also. So if they want to do it themselves, then they'll have to interact with the insurance companies. They have to take care of the you know, what we did was we didn't stop at the product. The product engineering is very limited there. So then we engineered the service. Then we engineered the service delivery. Then we kind of made sure that, you know, they don't need to go anywhere. They use our application. They, as in their service centers, use our application. And it's real time settlement of claims, et cetera, et cetera. So we continuously keep adding value to their customers and their dealers, whoever falls in the value chain. So they don't want to disturb that. And then suddenly it becomes so big, you know, so we've underrated 15 million of their customers. They don't want us to, I mean, they want us to take care of them for the next five to 10 years. So then even if they want to, they don't want to let go of us.

Dhruv Sharma: Interesting. What's your process for picking new industries to break into?

Tanay Goel - Garantie: Yeah. So this is interesting. You know, we started with automobiles, especially two wheelers, and we engineered that part. So we work with likes of Honda, Suzuki. So between them, 40% of the two wheelers in India today, you see on the roads might have been underwritten by us. So when we reached that scale, now for me, the next leg of expansion for two wheelers would have been going after a hero or a TVS. So either of them, if I get into my fold, we'll become the largest player in this space in the country. So, but it takes time. So while all of that is in the pipeline, you know, we say, okay, where can we take this expertise to? So that's when two years back, we entered mobile space. In the mobile space, we have a few competitors who are focused on retailers and distributors and creating value for them. And I realized that that's not the play. You know, you got to add value to the OEMs. And we bought the same playbook that we played into the mobile space. So that's, so for example, right now Vivo is one of our players who works exclusively with us. And we was the largest smartphone OEM company in the country today. They control 24% of the business, smartphone market. We are on route to probably build the largest OEM driven production plan business in the country ever done with that. So it takes time.

Utsav Somani: And this is something like the AppleCare plus plan that Apple has.

Tanay Goel - Garantie: Yeah, absolutely. So Apple has this global relationship with AIG. So Chinese smartphone manufacturers were not focused out there. So they were busy capturing the market share. And they did it very well. You know, they control 65-70% of the smartphone market.

Utsav Somani: Across categories, like mobile phones might be different, cars might be different, bikes might be different. How many, like, suppose 100 people buy like a mobile phone or 100 people buy a car, how many of them actually end up getting an extended warranty or a production plan?

Tanay Goel - Garantie: I can't comment on cars, etc. I think the penetration is going to be much higher. We don't operate in that space. We operate in two wheelers and mobile phone. It's highly, highly retail. And that's where we see value. So in the two wheeler space, 70% penetration we have on these products. In the mobile space today, we are operating at 3-4%. You know, at 3-4%, Indian market is around maybe 2000 crores.

Dhruv Sharma: Do people get cover mostly for like high-end phones or across the board?

Tanay Goel - Garantie: No, it's pretty interesting. So we thought that, you know, high-end customers, net promoter scores and brand cohesiveness is going to be much higher. And there'll be much, much higher value to be extracted there. But with these guys, they have very deep penetration in the lower segment of users. And for them, it's a far bigger problem if the phone breaks. So we see adaptability. To my surprise, adaptability is pretty high down there also.

Utsav Somani: And I think with the, I mean, as a final question, just wanted to ask you, like, I mean, you're running this bootstrap, you built a solid scale business across industries as well. And I think in this world of AI, like, do you think about that, integrating that? What parts of your business is it going to impact most? Like are you spending time thinking about that application?

Tanay Goel - Garantie: We're just not thinking, we're doing. So, you know, just to be a sense, we acquired today 400 odd thousand customers a month. My total team size on the ground is just 20 people.

Tanay Goel - Garantie: Yes. So a lot is done on the technology side. So we have real time. For the first time in the country, we were approving claims without manual intervention, without human intervention. So our system, the biggest trend that we have is like huge data of customers that we have now. The behavior, behavior analysis that we have now. So we made the systems learn, machines have started learning. Although I'm not too, I'm not a very big fan of they taking the decision, but I don't have a choice to let them do it now.

Dhruv Sharma: Yeah, I think it's key. That's what makes AI interesting to implement because, I mean, the models are trained in public internet data. And if you have, again, what's called public data.

Tanay Goel - Garantie: So it's a capital data. As I said, I'm not a very big fan of machines. Machines doing transactional data level work is okay. But when they start making decisions, I personally don't like it. But I don't have a choice right now. You know, I'm just falling in sync with how the industry is evolving. So we are very high-end technology team, which operates there.

Utsav Somani: Interesting. And you're claiming 40% cost savings through automation. So what specifically gets cheaper?

Tanay Goel - Garantie: Operating expenses on manpower. You know, you don't need offices, large offices, you know, because typically you need large offices to make manpower shift down there. So, you know, your administrative expenses goes down, rental goes down, your salary expenses goes down, all of that goes down. So how we build businesses, we invest in CapEx, but we don't invest in OpEx. So we make sure operationally, the business has to make money on every transaction that we do.

Utsav Somani: All right, Tane, thank you so much for coming on the show.

Tanay Goel - Garantie: Thank you for inviting me on the show.

Utsav Somani: All right, Lisbeth.

Saravana Maruthamuthu - optoML: Hi, what's up? Nice to meet you and nice to connect with you.

Utsav Somani: Awesome. Are you joining us from Singapore?

Saravana Maruthamuthu - optoML: I don't know. I'm in Chennai. I just came for a customer, you know, kind of, you know, customer visit. So just here. I'm not an office in Bangalore. So we have an office also in Bangalore. So that's our R&D center. So I'm in Chennai now.

Utsav Somani: So, yeah. So, I mean, why don't you explain what OptoML does for our listeners?

Saravana Maruthamuthu - optoML: Yeah. So OptoML, we are a low power AI inferencing company. So we build the chips, which are like, you know, 20x energy efficient than the conventional compute. And this is mainly used for all the inferencing. So it can be edge, like you have a drone where you want to run AI capabilities on it, or a data center. So that's where the unique USB that we bring in. And then we bring in to a new technology called analog in memory compute, where it's like, you know, the compute resides in the memory, because that's the major bottleneck of AI compute now. Conventional compute is mainly like, you have a lot of arithmetic going on, a lot of general processing going on. Now, it's all about data movement, right? So there's large data movement, large models are running. And so that is where the paradigm shift in the compute is needed. So that's what we are working towards.

Utsav Somani: So just to make the technology more accessible, what is analog in memory compute? Like you're basically doing compute in the memory, but if you're already building faster and faster GPUs, will this matter? And will this efficiency advantage that you mentioned last for long?

Saravana Maruthamuthu - optoML: Yeah, I mean, you can really think of the brain as a very good analogy, right? So it's like, I mean, I just want to little bit take kind of inspiration from neuromorphic computing to explain this concept. So basically, if you look into our brain, it's a very efficient system. It takes around 20 watts of power to process all the information. And there is no like a memory in a brain, right? So it's all about neurons connecting. That's how the deep, I mean, if you take an AI model, it's all about deep neural networks, neural networks, the neurons are connected to each other, and the neurons fire. And in brain, it's an electrochemical reaction where the neurons are connected and they fire, right? So that's how the, what is it, intelligence perception comes as a kid. So you learn things, you perceive, and then the neurons connect. And similar approach is what we, you know, is kind of very important for the AI compute now, because you want to build something that is more like, you know, memory-like, like a brain-like kind of thing where the system is able to do all the computing memory, like, you know, it's all about interconnects and how it's all about matrix multiplication, right? So you multiply large values of data, but if you do it in digital, right, even if you have faster GPUs, you're going to burn more silicon area and more power in the end. So that's where, you know, we are also able to get like 10x area efficiency to our approach. So all these things come together, helps you in latency, in operational cost, and also like, you know, the efficiency, and also the cost of the chip goes down. So these are the major advantages that the customer gets as a ROI for this innovation.

Dhruv Sharma: And you also, you said inference, so your focus is a lot more on inference and not training, and therefore the digital domain is separate, and analog is a great modality for inference, especially on-device inference.

Saravana Maruthamuthu - optoML: Yeah, true. So there are two things to it, right? So basically, we don't say that we do all analog. I mean, of course, if you see, you know, the inference, right? Inference is like, you train the model once, and then you kind of put into inference, and then it runs over and over. Take it, you know, any LLMs that are like, you know, 70 billion parameters of, I mean, currently we have 1 trillion parameters LLMs. So the key is, like, we build in a heterogeneous compute system on chip, that's where we uniqueness we bring in. So the digital part of it, where you want to do other, I mean, other aspects of compute, for example, you want to normalize the data, massage the data, get the results out. So that's all done in the digital domain. Only the, you know, the compute heavy part, that's matrix multiplication is done in the, you know.

Dhruv Sharma: Can you help us understand the trade-offs between these two? I mean, if one is more energy intensive, it obviously, maybe it generates more heat, and the other, or maybe there's precision loss in the other approach, et cetera.

Saravana Maruthamuthu - optoML: So as you rightly pointed out or hinted, right, so it's all about a trade-off between accuracy and power, right? But inherently, if you take deep neural networks, you don't need absolute accuracy. For example, if you do two into two, it may not, I mean, it can be 4.1, it may not be four, for example, right? So, and then when you do it at the matrix scale, you can always quantize the models. There is noise aware training that you can do to the model. And then you make sure what you're building is the full stack of it. I mean, the full stack to it is like, we kind of build till the reference solution to the customer. And of course, from there, the applications and other aspects of it is a long journey for the product. But what we are planning to do is like, we kind of bring in all the system aspects of noise, the accuracy part of it, right? Where you can do quantization, aware training, you can do noise aware training. So basically what you do, make the neural network learn the kind of the non-idealities of the analog world. And then you kind of, you know, make it more robust to it and make it viable to be deployed. At the same time, I mean, if you look into the digital side of it, right, so digital is always needed because all the data outside is digital, right? If you look into it, of course, there can be some directly data coming in from the sensor that can be analog, but predominantly all the data around is digital. So you need to have a digital interface around it. So for the general compute and also for pre and post processing of the data, because that's where the digital compute is having its power. And then you need to pack both of it together. And that's what we are doing. And then we are doing it in a, not in a very exotic technology where it needs maturity, the technology is available. So we are using our innovation to build memories that can be compatible with the digital and that is doing analog compute.

Dhruv Sharma: I wish we knew we were talking to basically the Dr. Tharoor of semiconductors. We would have brought our semiconductor dictionary and thesaurus along, but we haven't.

Utsav Somani: We spent 17 years in the semiconductor industry including Intel, Qualcomm and Continental. And I read something which I'd love to understand more in terms of timeline. So you've announced partnerships with, I mean, your supply chain, TSMC for Fab and KNs for OSAT. So you're truly playing out the whole India silicon, the India semiconductor mission story. But you announced something about a tape out with TSMC, the 12nm tape out is done. What does it mean if we were to explore this latest milestone in your journey? And how far are we to a fully working tested chip? Like what are the risks that lies, what are the timelines to come into market?

Saravana Maruthamuthu - optoML: Yeah. I mean, sorry for my jargon, I'll try to be like, no, no.

Dhruv Sharma: I mean, this is how we learn and our listeners learn.

Saravana Maruthamuthu - optoML: Sure. So, yeah. So about the milestone, when you want to build a chip, it's like, as I said, the ecosystem is very important, right? So you get the technology, it's a fabulous, we kind of work on a fabulous model, like other big companies work. So the fabulous model works that you get a technology from the Fab, right? And then, then you design your IP on the technology and you give it for fabrication. It's like, you know, you have, you are a kind of a furniture company, you're making a blueprint, you give it for the workhouse to get the design done. And that's what the Fab is doing for us, right? But getting the technology is a major thing, because you get to do, I mean, you are subjected to a lot of due diligence from the business side of it, and also from the technology side of it, how capable you are, and also how the, because the security of the technology is very important, what if that you contaminated. So that is a Fab aspect of it. So we kind of worked very hard to get the technology access. And from there, you need to have all the tools and flow to design it, because these tools are also, you know, it needs a lot of, what is it, legacy and multitude of tools to make the chip, you know, work on it, because you are working at 12 nanometer, there are a lot of masks, there are a lot of design tools around it. So all these have to be met. So this is the tool part of it.

Dhruv Sharma: On this, maybe I'll ask you to help us understand what photonics is. I think that'll be a great, you know, quick one.

Saravana Maruthamuthu - optoML: So the photonics part is like, basically, we kind of, you know, the data set, the data moment is very, very critical, right? If you're passing the data through copper, then there is a limitation of the wire, physical wire. So we want to change the media and then make the data flow or interconnect between them using optical optics. So that's where you kind of get an optical highway, where, you know, as I said, the data is a bottleneck. And then if you have an optical highway, it's like, you have like an express way where the, you know, copper does like Ohm's law get in the way, like resistance, etc.

Dhruv Sharma: Does that lead to a problem?

Saravana Maruthamuthu - optoML: Yeah, I mean, it's a speed of it, right? Because copper is subjected to a lot of parasitic. So I mean, there's a physical media like that, you can move it like, I mean, you can just think it as a byline broadband, right? You had the violent broadband, and then now we have the optical fiber. So you really see the difference in it, like how much throughput you get. Similarly, you can imagine through it. So it's like, it's limited by the physics of transmission media.

Utsav Somani: And your recent $1.8 million raise that you've announced, congrats on that milestone as well. Where does it take you? And I'm guessing your journey will require a lot more capital as well. And what are the milestones you're looking at? Yeah, what are the metrics that you think you need to hit in terms of scoring on the ground?

Saravana Maruthamuthu - optoML: Sure. So currently, the test ship, you know, has to be functional. And this is a very big milestone that we achieved within 11 months. And then it's a new team using our capital efficiently. So we believe in utilizing the capital very efficiently and work on it. So it will help us to get through like, you know, end of this year, and where we want to do another race where we want to go in for the full product ready that will be available in 2027 Q3. It's a very patient capital. And then the design cycles, each of the design cycle roughly takes around one year of time. So and this makes it I mean, we are in a fast execution cycle, we want to execute within six months of one iteration. So you know, from a beta version to an engineering sample. So that's where we want to be. And then like there are there are multiple market segments that we are also targeting. Edge compute is what we're targeting first and then moving towards hyperscaler. So that's our mission. Hyperscalers are more targeted towards 2028, 2029.

Utsav Somani: Through you, we were discussing the news that we saw, right? The Microns new plant.

Dhruv Sharma: Oh, yes. What was that inauguration about?

Saravana Maruthamuthu - optoML: Yeah. So I think Micron, you know, like is also kind of I mean, the memory is very key, right? So they are I mean, inaugurating a fab in India. To be honest, I'd like not like the details of it. But what I know is like, I mean, memory demand for all the you know, all the computer RAMs, SRAMs, I mean, the DDRs and HPMs are going up very high. And like some of the I mean, that is also creating a lot of demand within the you know, within the automotive sector or consumer sector. So I think that is one of the major things in India. And like that's like basically I mean, the ecosystem is building up in India. So that's very nice news for us. And also like we want to be part of the ecosystem. So that's what we're trying to do.

Dhruv Sharma: Yeah. Any final closing question, Dhruv? Yes, we do have time for one, right? Yeah. So, Saruna, you know, when the AI Impact Summit was happening, Serum had announced this partnership with Nokia. And then, you know, everyone came warmed up to the fact that maybe for India, on-device inference might actually be the way to go. Do you have a point of view on that? And why that entire round trip from the cloud in terms of like inference and cloud is maybe not so efficient for every nook and corner of our country?

Saravana Maruthamuthu - optoML: I think, see, some of the models, if you really look into what is it like useful AI, right? I mean, you don't need LLMs, large models for everything. For example, now we are working with Vision QA, intelligent drones, surveillance cameras, where you want to do a kind of inference, which is like focused on a particular application. So you don't need to rely on the cloud if you have the compute power on edge. And then you can do a lot of heavy lifting on the device itself, and then pass the data to the cloud for, you know, for the database.

Dhruv Sharma: One very natural question is what if the weights change or what if you want to switch models, then how do you, can you, can you do that?

Saravana Maruthamuthu - optoML: Yeah, I mean, like, it's like we build our technology on SRAM. So basically, it's like writing into a memory, right? So we can write the weights into the memory. So we have a programmability and we just have an...

Dhruv Sharma: So you can reprogram.

Saravana Maruthamuthu - optoML: Yeah, it's like, you have connect to the, to the, you know, like to the device through the internet.

Utsav Somani: Yeah, I think we have time for one more question. So I think I want to go more philosophical and zoom out a little bit. India is making a big push in the ISM 2.0 mission, right? And they want 50 fabulous semiconductor companies by 2030. Where do you think we're at in terms of this journey of achieving that target from a founder perspective?

Saravana Maruthamuthu - optoML: Yeah, I think it's a very nice question, because I see that it's like only when the fabless companies develop, I mean, I see like a lot of companies are coming up and they're working on very interesting problems from RF, from power to microcontrollers and AI compute. And I think that that is the key driver for the volumes for the fab and the OSAT. So I think that is one of the key essential components that we need to focus on. And I mean, it's not like a one company plates, it's a multi-company play. And then if you look into the components that are, you need to source multiple components for an application. So I think this is very interesting. And I think, I mean, like we are delayed, but I think, yeah, we are, we are on it now. So I think that's what I see that for India.

Utsav Somani: All right. Thank you so much for joining us on the show.

Saravana Maruthamuthu - optoML: Thanks a lot. My pleasure.

Nayrhit B - Gushwork: Hey, how are you?

Utsav Somani: Loving your LinkedIn post, by the way, you're truly building in public. And we have a very common friend who was excited to tune in. I hope he's tuning in. He said, Karan Kumar, I don't know if you know him.

Nayrhit B - Gushwork: Yeah, I know. He was the one who introduced us to Lightspeed. So he's pretty much responsible for whatever it is.

Utsav Somani: What a shout out. Of course, stuck in Dubai right now. So I hope, I mean, everyone who's there or affected by the war, I think is staying safe. And hopefully this passes us. Let's introduce Kashwag to our listeners. What do you do? What does the business do?

Nayrhit B - Gushwork: So we pretty much help small and medium businesses get more customers and get more leads from this new emerging marketing channel that is AI search engines. Pretty much like every time a platform shift has happened, there has emerged a disruption in the way marketing budget is allocated across channels, be it when yellow pages came in, Google came in, Facebook came in, et cetera. So now we have perhaps the biggest platform shift that has ever happened in a digital world. And we're seeing that pretty much as we are building this as well, and how the tailwind is affecting the ROI that our customers see out of our product as well. So that's pretty much what we do.

Utsav Somani: World is moving from SEO to GEO. And is GEO the same as AEO? Or what's the terminology that's been popularized?

Nayrhit B - Gushwork: I think the jury is still out on what's going to be the terminology that sticks. But pretty much, I mean, the fundamentals and the first principles of marketing is still not very different. And that not just applies to search, a traditional search or AEO. It also applies to Instagram, LinkedIn, et cetera, as well, in my opinion, any organic content channel, where every time you create something on your native platforms, you need to add value and answer someone's question. And then you have to figure out a way for other people to promote your answers, which is called backlinks in this universe. But pretty much those first principles still hold true. The form factors and the specificities and the nuances have gotten more complex and more algorithmic in nature than it used to be, how manual it was. But to answer your question, Utsav, now, at least the popular term is AEO, as popular as by HubSpot right now. But yeah, LLM, SEO, AEO, GEO, et cetera, whatever you want to call it. We are not using any terms right now. We just say get leads from AI search engines. And that also encompasses traditional search engines like Google, which is introducing AI overviews. And in about a year, we'll stop seeing the links pop up there. So everything will be AI search. There's nothing called traditional search anymore.

Dhruv Sharma: So that old world of search engines, you'd query something and you'd get a set of pages with links. Now you ask a question, you get a very synthesized response. So maybe you could just help us understand how the architectures of the two systems are different, right? Like search engines pretty much used to be like retrieval systems. AI is just different. But more, I mean, from an actionable standpoint, people who care about brand visibility, how should they be thinking about this? How should they be refactoring their websites, their content and so on?

Nayrhit B - Gushwork: Yeah. Dhruv, I think sends two questions there. Question one is, how is the architecture, the technicals of this ecosystem different?

Dhruv Sharma: The keyword world and the prompt world.

Nayrhit B - Gushwork: So I think very interestingly, I'll try to not go super technical, but to answer your question, Dhruv. So fundamentally, what's happening is, let's say earlier, you needed to, let's take an example, you needed to, let's say, plan a trip to Goa. Now to do that, you would yourself search maybe flights to Goa and MMT would prop up your book, your flights. Then you would maybe go to Airbnb, look at a villa, and then book that flight, book that villa. Then there would be a third search maybe a month down the line, looking at like things to do in Goa. So Little Black Books website would pop up or some blogging website would pop up. Now, instead of that, now this broken search from the first conversation to actually the experience happening, that's pretty much like a long drawn month to two month process. And here is where a lot of drop offs happen. But now what you're asking is, hey, can you build me a 10 day itinerary for Goa? Please include the flight details and stuff like that. Imagine you are like a specialized travel blogger or something, and then you would have like an elaborate prompt on it. So now what's happening? Does this mean that the search is not happening? It's still happening. But the agents on the backend are being spun up by these models. And they are in turn breaking down your question into multiple search terms, and actually searching the pretty much the same results that you used to search for online. So what's happening is, I call this TAM explosion. So your engagement time with ChatGPT or Claude or search is actually increasing. And AI search is not actually competing with traditional search, but is actually competing with word of mouth, referrals and socials. So all the questions that you would ask your friends, all the questions that you would ask, not ask for social stigma, all the questions that you would consume on Instagram are actually now coming to ChatGPT and Claude. Now that's why what I feel is there is a massive explosion of searches happening. In fact, very recently Sundar Pichai had a podcast with Verge. He mentioned 15% of questions asked on Google or Gemini every day is never before seen in the history of mankind. So very much the number of questions we are asking on the internet is growing 15% every day, the distinct queries. So that itself tells you, these questions, does that mean that they never existed? I don't think so. They actually existed, but we never asked either because of social stigma or we asked our friends, it was never digitized in the first place. So to answer your question, first part of the question, the searches are still happening and much, much larger volumes of searches are happening. But this time it's AI agents which are searching behind the scenes and synthesizing the answers and getting the answer in front of you essentially. And now what's the fundamental difference on how the AI agents are operating? And there is obviously a lot of tactics. I'm not going into the tactical stuff, but if you look at this Can we share that actually?

Utsav Somani: Like if somebody, like a small medium business wants to get discovered by buyers on ChatGPT, what would Gushwork do for them? I think that might be interesting for listeners.

Nayrhit B - Gushwork: Definitely. So one of the most important things that where we start our analysis and where the product starts happening, we call this the memory agent. And what the memory agent does is it scans your website and it takes a lot of input from the customer and this part is very critical. So it's actually manual. So the way we design this agent is after collecting a lot of the information from customers. So what are the services they offer? What are the products they offer? What's the catalog, which is digitized or non-digitized? What are the areas they cater to? What's their ICP, et cetera? Using this information, we feed it into what we call memory agent. Now this memory agent is used to ground the results of all the subsequent agents that are about to come. The second agent is what we call the research agent. The research agent is going to try and hit multiple databases like SEMrush APIs, Ahrefs APIs, Google Keyword Planner APIs. Hopefully, ChatGPT will release something to the equivalent of Google Keyword Planner. So we'll start hitting those APIs as well. To guess what are the long tail questions your buyers could be asking on the web, but there's not enough data on that. So there's not enough content created on the web. And so to give you an example, let's say earlier the question would be best CRMs for SMBs. Now the question is, hey, imagine you are Jason Lemkin, how would you calculate sales quota and what's the best CRM to do that? Now the question is this. If you look at any traditional search tools or any SEO tactics, they would ignore the second question because the volumes are too low and it would not interest the market. But these long tails is where the play is getting very interesting because this is where when you write an answer, there's pretty much no answer on the web which has answered this specific question. So that's where this third agent comes in, which is trying to find out the 100 to 200 queries which are being asked about your business and trying to answer them by tapping onto your internal as well as external data sources. And thus these answers get created, answering each of these questions. Now, does it mean like all of the pages and all of the answers we create gets cited? It's not the case. We create about 100 to 200 assets in one shot and we continuously keep updating them based on the signals and the mentioned data that we get from ChargeGBD Gemini. So for each of these queries, we start seeing if your brand is getting featured or not. About five to ten percent of your pages start getting cited within the first 60 days of making them live. And once they get cited, the buyer visits your website and they become a lead for you. They fill up a form or something.

Utsav Somani: And you've publicly written about your ARR reaching 1.5 million in three months. What does the payback look like? What is the CAC? And also you've publicly said that you want to hit three, three and a half million in three months. So that's crazy growth.

Nayrhit B - Gushwork: Yeah, thanks a lot. So look, our payback hovers anywhere between three months to nine months. Now, we're trying to get that down further and which is why press release and all of that as well. So we get a lot of inbound search of leads as well. So, so far, we have been pretty cold, outbound centric kind of approach. Now, we are starting to see like three months payback as well. And the second question that you asked is like, what's the CAC? Our pricing anywhere hovers between 10K annual contract values to about 25K annual contract values. And our CACs are about hovers between 1000 to 3000 dollars.

Utsav Somani: And do you, I mean, do all models react the same way, like agents from all models? Like, I mean, perplexity was, I think, the one that started off this AI search wave, right? And now I think they're still finding their mojo. But like, does Chantjypri search, which came later, or like Claude internet search works differently?

Nayrhit B - Gushwork: Good question. I still, I wish I had a better answer for you. But Claude is still not giving UTMs. So it's hard to attribute traffic coming from.

Dhruv Sharma: How is the search traffic distributed across the, you know, most popular ones?

Nayrhit B - Gushwork: Yeah, very good question. And the data sort of keeps surprising me every month, because it depends on the industry that we are going after and who we are catering to. So if you take a typical, let's say, very traditional industry, right? Let's say manufacturing or industrial distributor with the US or let's say a commercial janitorial service or something like that. Now, these types of customers see a very heavy distribution on traditional search as well as co-pilots and a little bit of Chantjypri as well. And when I say a little bit, typically we have, we measure two things. One is what is the bot traffic to human traffic? And then what is the lead share as well? So that bot traffic is approximately, let's say 20% skewed towards AI search and 80% towards traditional search industry engines for these traditional industries. Now the same thing you apply to a SaaS, let's say something which is catering to dev tooling market. Now there you see a very heavy share of AI search engines on both the bots being sent as well as the lead share that you are seeing as well. So that's how the, it honestly like differs based on industries.

Dhruv Sharma: Natural question is, you know, so the, this, I mean, traditional search would reward optimization, I guess, like keywords, just like, I mean, that's exactly, you would optimize your copy for it. What do you optimize your copy for now? What do AI search, what does AI search reward? Because, you know, it's like if five of my friends were sitting in a row and we use the exact same prompt on the exact same model, my guess is we'd all get five different responses.

Nayrhit B - Gushwork: Good question Dhruv. And I think that's, that's actually the way we optimize content for keywords. I think that itself is a bit of a hacky SEO. In fact, the great SEO guys actually already have been doing this for quite some time and we call it intent instead of keyword. So let's say the same semantically, the same keyword, which looks semantically similar for all three of us. Let's say, let's say I type in Apple. Now Apple could mean Apple, the stock price, Apple, the device or Apple, the fruit. Now, what is Tragibity going to show? It's going to depend on what your past searches have been. Now, if you have a little kid at home, it's likely going to show you Apple, the fruit, fruit, and all your queries are around grooming a kid or something like that. Now, this is what we call targeting intent as opposed to targeting keyword. Now, earlier the thought process was, hey, Utsav would maybe publish a podcast, a newsletter or a podcast transcript on his website and then try to like put some keywords in it or paraphrase some of the words in it to get it to rank for those keywords. And it would rank in those cases. But now what's happening is the fundamental topic and the outline and the content is decided based on what is the question being asked. So it's not that you create the content first and then figure out where do I get this ranked. Instead of that, you actually find the question that needs to be answered by the buyer. Then from that, you arrive at what is being cited right now. And based on what is being cited right now, you need to create an information gain on what are the citations that are coming up right now. So you pretty much need to be the best answer in the world for that particular question that is being asked by your buyer.

Utsav Somani: And how do you build custom tooling within GushWorks to make all of this process easier? Because you pivoted or sort of evolved from AI plus human company, right? More focus on AI search visibility now.

Nayrhit B - Gushwork: Yeah, exactly Utsav. So we actually built like six agents by mapping out these processes. These are the exact processes we were doing manually for almost a year and a half before we figured out like what is it that we need to automate exactly.

Dhruv Sharma: You guys call this AI feed, right?

Nayrhit B - Gushwork: Yeah, we call it AI feed. So it's sort of like a parallel to RSS feeds when Google came about. But AI feed is sort of a way to create content on website at a very high velocity. So typically, like if you notice like there is always a lot of traditional SEO market has frowned upon the fact that you're creating like content at a high velocity. But it's also the fact of the matter is that LLM's ability to consume text is also much higher than humans. Now, Webflow, WordPress or any traditional CMS is designed to produce content manually and manage them manually. Now, they're traditionally like people have not published more than 15, 30, 100 blogs a month. This is the first time we are going into a case where the content creation velocity is as much as Instagram's or Dick Denton's. So that's why you need to think of websites as an asset where you actually create everything that your brand represents, all the questions that your buyers have answered them. And don't think of it as a blog 2.0, but rather think of it as a news feed, which are consistently streaming to LLM searches in a fresh, signal-led manner that allows LLMs to map out what you are, who you are, what are your strengths, what are your weaknesses, what's your pricing, what's your case study, what are your testimonials? So earlier, we need to start getting rid of the fact that a human is going to come and read all through the content and you need to have a visually appealing page. So that's the thought process, essentially, going on.

Utsav Somani: Reading content, machines doing commerce, machines doing everything, basically. So as a final question, congrats on the 9 million raise as well. Three years out, crystal ball glazing. Where does this industry head?

Nayrhit B - Gushwork: Good question. I think one of the first things that...

Utsav Somani: I know, three years is a very long time. Three years back, we didn't even have Caterpillar. Let's reframe that. Let's do one year, actually.

Nayrhit B - Gushwork: Good question. Again, it's changing every couple of months as well. And these are what seemed impossible earlier is seeming possible every passing month. I think two things will definitely happen. One is we will see some sort of a connection between websites and LLM crawlers directly. Players such as ourselves, as well as traditional CMS companies, will try to figure out a way to expose the architecture of a website within UI of Gemini and ChargeGBT. That actually will allow us to fulfill the loop within the UI of Gemini itself. So today, people are getting worried about not visiting websites, but maybe they don't need to. Websites are just databases, and the UI could be Gemini or Cloud, essentially. So being able to fill up a form or being able to schedule a demo or signing up for a product, all of these workflows can be done if there's a sort of MCP-like layer between websites and LLMs. So that's one thought process which I think will be a reality sooner than later. The second thing is people will start thinking of this channel more as a brand marketing channel than a direct marketing channel. So earlier, there used to be like a clear attribution, hey, this is the query I'm featuring for, this is where I'm getting a lead for, this is the CAC of my SEO, this is the attribution. But I think it's going to be like pretty much table stakes. It's going to be very similar to billboard advertising, where you have impressions, and then you get leads.

Dhruv Sharma: I'm sorry to interrupt, but if you were to give like three points of well-meaning advice to founders, a conversation they should be having with their marketing heads tomorrow morning, what would it be?

Nayrhit B - Gushwork: Don't ignore these channels. A channel arbitrage is finished like radically fast. So you need to like move ahead. And there is an arbitrage for every channel which you need to capitalize. Second is, don't be paranoid about trying to balance marketing for humans and marketing for AI. There's a lot of things that are temporal in nature. There are a lot of things which seem tactical today and illogical today. They might actually be the future itself. So these are the two big things. It's not three. So I think the biggest problem we face is principally...

Dhruv Sharma: Three is made up, by the way, all of that McKinsey stuff. It doesn't always have to be three. Two is good.

Nayrhit B - Gushwork: Four is good. Yeah. No, I think it's largely a principal agent problem or like an innovator's dilemma. Most companies are facing that, hey, I'm getting so much traffic already. Like, what do I do? How do I integrate you guys? Etc. So that's my two cents on adapting to this.

Dhruv Sharma: Founders get to work.

Utsav Somani: All right. That's a solid note to end this show. Thank you so much for coming.

Speaker 7: Pleasure meeting you.

Utsav Somani: We're taking Wednesday off on account of Holi and we'll be back on Friday at 4 p.m. Thank you so much. Have a wonderful week ahead.

Dhruv Sharma: Wish you all a very...