Full Transcript

Dhruv Sharma: Hey there listeners, it's the middle of the week. I hope this one is going well for you. We're going to dive straight into it. Our first guest, Rohan Parker, is building a company called Spintly. And the first time I read about what the company does, I was like, man, Rohan can actually be a technical consultant for the next Mission Impossible movie. But with that, let's bring him on the show. Hey, Rohan.

Rohin Parkar (Co-Founder & CEO, Spintly): Hey, hi Dhruv. And how's it's up? Nice to meet you.

Utsav Somani: Good to have you on the show. So you're coming fresh out of a raise, 8 million from Axel. Let's introduce what Spintly is doing for our listeners.

Rohin Parkar (Co-Founder & CEO, Spintly): Yeah, Spintly is an access control technology company. We enable digital access solutions for large real estate establishments, corporate offices, co-working spaces. But instead of using your access card, you can use your phone for access, streamline visitor management using QR code based access, really getting to real estate space. But core technology is around wireless, mesh, IoT. That's where our core stack is, technology stack is.

Utsav Somani: Can you go one level deeper? I read something about a Bluetooth low energy mesh network as well, which enables this.

Rohin Parkar (Co-Founder & CEO, Spintly): Yeah, yeah. So the main differentiator of our technology in the market is you build a whole solution around Bluetooth mesh technology. Traditionally, these systems or any building automation system usually has a wired infrastructure behind it. So every sensor, like in this case, it's a reader which is talking to your card, is connected to a controller which is sitting somewhere in the building. So imagine a large corporate tech park. You probably have like 1000 doors, 500 doors in the building. Every door has this reader mounted and you're just tapping your card and walking in. But what really goes behind it is there is a wire going from every door back to a building automation room. So you can imagine the amount of wiring and cabling which goes into these systems. So we completely eliminated the controllers and the wiring and used IoT, basically wireless mesh networking technology to create a new architecture for access control. So that's our innovation in this space.

Utsav Somani: All of this has existed for decades, right? I mean, Bluetooth low energy has been around for a long, long time. You've looked at this access control problem with a fresh lens. Like what enabled this? I think you spent time with Nokia and Broadcom for this.

Rohin Parkar (Co-Founder & CEO, Spintly): Yeah, excellent. I mean, yeah, it's all about timing. You said it right. You could have built it before, but there were no standards around it before. There was a new standard called Bluetooth Mesh Standard, which got created by Bluetooth SIG, which is a special interest group in 2017 or 2016 timeframe. And that I used to work for Broadcom. So I was aware of that standard. So that's when we were bouncing on our ideas about smart home solutions using mesh technology or building automation solutions. And we stumbled upon access control. And at the same time, phone based access was just picking up, right? So people are talking about using phones for door access and things like that. So we kind of coupled this idea saying, hey, people are going to use phones or any way for access. Why not use the Bluetooth to also create the networking infrastructure behind it? So that was and the standard was there. So it was also timing. Right. So that's that's how we started. Yeah.

Dhruv Sharma: Rohan, is it that, you know, each of your devices installed on the doors act as nodes and sort of relay signals from one to the next and and eventually, so you've decentralized all of this in a sense.

Rohin Parkar (Co-Founder & CEO, Spintly): Exactly. So we call it as a distributed architecture for access control or building automation where the intelligence in the building is kind of spread out now, you know, at a broader sense where every reader actually has all the information needed to grant access. Usually it is a centralized architecture. So it's a distributed, decentralized architecture, you can call it.

Dhruv Sharma: How big is the largest facility that you guys that you've installed these readers in?

Rohin Parkar (Co-Founder & CEO, Spintly): I think we have about 200, 250 doors in one of the buildings, 200 doors in a few other buildings. These are large buildings in Mumbai or Bangalore, like either corporate tech parks, co-working spaces. So there's a lot of traffic. One of the buildings has about 100,000 people walking in pretty much every day.

Utsav Somani: And the pitch to them is cost advantage in terms of saving wiring costs or do you pitch like something else as well?

Rohin Parkar (Co-Founder & CEO, Spintly): No, we pitch actually the experience and the solution, right, where you are, if you have to have 100,000 people walking in every day, you can imagine sometimes the lines which get created near the front desk for visitors, employees. And usually these facilities have these turnstiles, right, that you have to tap your card and walk in and visitors usually line up because they don't have the access, right. So we actually pitch that, we streamline your access and we digitize it from end to end. So we have complete.

Utsav Somani: And by NFC, I mean, by 2017, I think most of the phones would have had NFC reading capabilities as well. But because of the visitors, I think you might have not allowed NFC to be the dominant mechanism for this.

Rohin Parkar (Co-Founder & CEO, Spintly): NFC, the problem with NFC is not a problem. Apple doesn't allow usage of NFC without using wallet. So we are also partnered with Apple so we can offer NFC solutions to Apple wallet. But for Android, we do offer NFC, but visitors don't download an app. So they need another way. So we have a QR code reader, which we send QR codes. And these are dynamic QR codes which are encrypted, secure. So you cannot just take a screenshot and share with a friend. So we have also done innovation on the QR code front. So our vision, our goal is to just streamline movement of people into corporate offices and office spaces.

Dhruv Sharma: I was going to ask you if you pitched the Louvre, they had a painting stolen from them last year or maybe it was this year.

Rohin Parkar (Co-Founder & CEO, Spintly): I mean, yeah, you know.

Dhruv Sharma: Although the most serious question to ask there, Rohan, is what for the vulnerabilities with that original home run sort of system that you guys identified and therefore your solution is markedly better.

Rohin Parkar (Co-Founder & CEO, Spintly): Actually, you'll be surprised the protocol which was used for the wired system is called Vegan Protocol. It is actually an unencrypted open protocol and it is completely vulnerable. You can actually sniff the signal going on the wires and actually read the credentials information, right. But because it was on a wire, nobody thinks it is actually vulnerable, but it is actually pretty unsafe. With wireless mesh, it is fully encrypted, the communication is encrypted. So there was a new wired standard which also is launched for OSDP recently, which is an encrypted wired protocol. So to fix the problem of this, we went a step ahead and said that we can do encrypted wireless protocol. So in fact, we are trying to define this as a standard now. We are working with some bodies to define this as an open standard for access control.

Utsav Somani: And where does this 8 million raise take you? I believe you have 500 Indian customers, but you're also spending time in the US where you're dialling in from now. So is the focus increasingly going to be international markets?

Rohin Parkar (Co-Founder & CEO, Spintly): Yeah, India obviously is a growth market for us, given the current pace with India is growing, but you're expanding beyond India and US, Middle East. We started our office in Middle East, but unfortunately things happen right now in Middle East, so things are a little bit jittery. But that's a big region. Southeast Asia is another region we want to expand.

Dhruv Sharma: Have you mapped out the different opportunities that's growing, like access control is required everywhere, like hospitals, hotels, office complexes, so on and so forth?

Rohin Parkar (Co-Founder & CEO, Spintly): Yeah, definitely. I think our focus has been commercial offices because that's where there's direct ROI, which they see. But other than that, there's huge opportunities in public infrastructure, whether it is airports and even government facilities where access and security is a huge deal. And hospitals, as you said, and also hospitality is a big segment where, you know, you use actually smart locks. Our technology also goes into smart locks. We partner with lock companies. We embed our hardware and our firmware into smart locks. So there are various opportunities. Automobiles is a long term opportunity, keyless access for cars and things like that. So the technology has quite a few things.

Dhruv Sharma: Also asking this question for other founders who will at some point have similar decisions to make. I mean, you know, the world is your oyster, but you've got to pick wisely, pick well and sequence your entire journey. So talk to us a little bit about how you've made those decisions.

Rohin Parkar (Co-Founder & CEO, Spintly): Yeah, I mean, we had to actually narrow down it. So when we started, we were a smart home company. We were actually doing smart lighting for homes and, you know, smart locks and pretty much everything we wanted to do. And smart locks kind of evolved into access controls. And and we kind of gave up everything else, like we gave up all the lighting and whatever we were doing and just focused on access control. Yeah, so I think focus is really important for a startup where you can use the capital really well. Even today, we keep thinking about, hey, what not to do rather than what more to do.

Utsav Somani: But when do you I mean, what was that moment in time where you landed up on that insight that you wanted to shift from smart homes to this? Like, was it a customer feedback? Was it a particular use case that was increasingly becoming evident or was it just technology shifts?

Rohin Parkar (Co-Founder & CEO, Spintly): No, it was definitely customer feedback and market dynamics. We saw that smart home market is very crowded and it is a consumer market, right? So I mean, not a bad market to be in, but Google and Apple kind of control the space because they control the phone ecosystem, the app ecosystem. Right. And we saw that access control is very slow to move. There's not much innovation happening in that space. And there is direct ROI. Consumers don't like to pay subscription fees. Right. If I ask you, hey, if you turn on your lights, you have to pay a monthly fee. You're not going to pay that. But enterprise is ready to pay up subscription fee for solving their problems. So that's where we said, OK, this is much better. We can make a lot more money here than going.

Utsav Somani: How big is this market? Like if you were to just focus on one particular vertical, like how big is this market, like the office going segment?

Rohin Parkar (Co-Founder & CEO, Spintly): So physical security market globally is is about 100 billion dollars, 100, 100 plus billion dollar market. Our biggest competitor here today makes about one point eight billion dollars in annual revenues. Right. And that's incumbent. Right. And there are many other larger players who make similar scale of revenue. So that's how big the market is. And today with we are also incorporating facial recognition systems with that, that market even becomes bigger where biometrics and AI is having a big impact on security space.

Utsav Somani: And will this be relying on existing CCTV infrastructure or are you planning on developing or licensing some hardware from other players?

Rohin Parkar (Co-Founder & CEO, Spintly): In fact, both. Right. So there's a lot the bigger opportunity is the brownfield where how can you convert every camera into a smart camera today using Edge AI, where you can run the analytics at the edge and make decisions. So without having them to invest heavily into new infrastructure, that's that's the big opportunity.

Dhruv Sharma: Amazing. Do you guys have like a R&D lab sort of thing inside the company?

Rohin Parkar (Co-Founder & CEO, Spintly): Yeah. So we are fully vertically integrated. We develop all our hardware in-house. So we have R&D facility in Goa. I'm from I'm from Goa. Me and Malcolm are the co-founders. We both grew up in Goa, childhood friends, studied engineering together. So we have about 100 plus people in Goa. And then we have a sales office in Bangalore. And so we have two physical offices and one office here in the Bay Area.

Utsav Somani: And yeah, Dhruv.

Rohin Parkar (Co-Founder & CEO, Spintly): Well, I was just going to say you started your career at Bark, right? Yes, I was a scientific officer or a scientist at Bark. We were doing research on particle accelerators. That that's how I got into the RF engineering space. Right. And all my life, I've been an RF engineer developing wireless technologies and systems.

Utsav Somani: And what kind of other requests are Indian enterprises making in terms of I mean, we're getting like these big smart building market. I mean, these co-working spaces to like these big massive office towers, big campuses as well, GCCs of many big players coming to India and already established here. Are they making specific requests in this? Yeah.

Rohin Parkar (Co-Founder & CEO, Spintly): You know, the requests are huge, actually. I mean, the amount of requests we get is crazy that but that's what I was saying. What not to do it and what to do is what we have to choose. Like meeting rooms is a big booking is a big request from a lot of large co-working spaces, enterprises, because they monetize these spaces. So they want a software model within that because we can do it easily access control. Right. So you can access control every meeting room. Visitor management has been a big ask and that we actually developed it, rolled it out. And as cafeteria management, a lot of things which are tied to access control is what they want us to offer them. So it's a partner versus build kind of a dilemma we have to face. Do we partner with a meeting room booking software company and let them integrate into our APIs and SDKs? That's one approach we take. We take that. We prefer that approach many times because it's like joining forces and growing together. But many times customers like I just want one stop solution. I want a single pane of glass for managing my entire building, everything related to access. So time and attendance is another. They want to manage their employees time, leave management, like a sort of tiny HRMS kind of a solution. So there's a lot of things which can be built on top of access control.

Utsav Somani: You can become like an access control operating system of sorts for facility management with that sort of by coming to you. But what are the unit economics like? How do you generate revenue? Is it like a deploy one time and then annual subscription or is it pay per access or how do you charge your customers?

Rohin Parkar (Co-Founder & CEO, Spintly): I mean, we make money on hardware. So we do sell hardware first and that is we're building a channel partner network across the country, across the globe who will resell our hardware solutions. So we make decent margin on hardware, 40 percent gross margin on the hardware sales. And then it's a subscription based model where we charge for every door which we manage a base fee. And then for mobile credentials, there is a per user fee. And then visitor management is based on number of visitors, subscription fee, annual fee. So everything is usually charged on an annual basis, subscription fees. And is the hardware manufacturing localized at this point? Yeah, all the hardware is manufactured in India. It's a design in India, manufactured in India. We have a couple of manufacturers who we work with, contract manufactured in India, Gujarat and Hosur, Tamil Nadu currently also looking at a manufacturer in Gurgaon now. So so we kind of have working with three manufacturers today.

Utsav Somani: And what's next as a focus area for you?

Rohin Parkar (Co-Founder & CEO, Spintly): Next, I mean, growth, I mean, I'm an engineer by background. So constantly we have to go divert my attention from product to business. That's always we get pulled into product. But again, focus is how do we grow the company? How do we increase the top line and increase the partner network, channel partner network? That's the key. Because our competitors have a huge network of channel partners and distributors. That's their strength. How do we break that? That's the key for us today.

Utsav Somani: Yeah, I think that's it for me. Any final closing question from you?

Dhruv Sharma: That's it. We should let Rohan go back to securing more and more buildings.

Rohin Parkar (Co-Founder & CEO, Spintly): Thanks. Thanks, Utsav. Thanks, Dhruv. It was a pleasure. Thank you so much for coming on the show. Have a wonderful day ahead. Yeah, thank you. Bye bye. Take care.

Utsav Somani: All right, listeners, we're moving on to our next segment. We've got Sajo from Deccan AI joining us now. Sajo, welcome to the show.

Sajo Mathews (ML Lead, Deccan AI): Hi, thank you, Utsav. Good to be here.

Utsav Somani: Congrats on the recent 25 million announcement. But I think more interesting is your name. What's behind? What's the story behind Deccan AI?

Sajo Mathews (ML Lead, Deccan AI): Deccan AI? Oh, well, so Deccan AI, I think it was there were a lot of companies which have started up, you know, which kind of indicate their origin story. Not so many from India. So I think Rakesh, when he started the company, he really felt that, hey, we need to have an Indian representation which proudly talks about its origins. And because he started in Hyderabad, so it became Deccan. That time there was no Bangalore set up. But, you know, I've recently joined. So now I'm based out of Bangalore. So we kind of cover the entire Deccan plateau. So it becomes more real.

Utsav Somani: Amazing. And now that you've told us the story about the name, but also tell us about the company, what is Deccan AI doing?

Sajo Mathews (ML Lead, Deccan AI): Yeah, so Deccan works right now with Frontier Labs. Right. And we've been helping, you know, Frontier Labs with training data to build their models. So if you look at models, we actually need to have the algorithms, which is what the Frontier Labs themselves build. You need to have the infrastructure, which is where, you know, the NVIDIA and friends can come into play. And you also need data to be able to train those models well. And we kind of work on the data side of the business, where particularly in terms of post-training data. So, you know, that's after the Frontier Labs have trained their models on whatever is available on the Internet and stuff they have access to, you need to be able to align the models to become experts in certain domains, say, mathematics, science, medicine and so on. Or you need to be able to evaluate what are the places where the models are actually not doing well. So that's where they kind of rope in folks like us. We have a massive network of, you know, over a million freelancers based in India. And yeah, and these folks basically come from, you know, multiple social expertise. We have folks from IITs, PhDs, doctors, lawyers, all on our platform who help grade these models upwards or provide specific training data that they need to become better. So that's what we do today. And now we are looking to see how do we extend this into enterprises as well to help them build agents. So that's really what the fundraiser is about. Amazing.

Dhruv Sharma: So typically, how much time does the Frontier Lab have between, you know, the pre-training process finishing up and then the release date?

Sajo Mathews (ML Lead, Deccan AI): Oh, between pre-training and release date. Well, actually, Frontier Labs are always working on iteratively, right? So it's not like there is, you know, a fixed date per se, but you're always looking for understanding what are the gaps. And then every time there's a gap, there is. So there is obviously a specific window that the labs themselves decide on. This is the date at which there is a cutoff date for knowledge. And then you kind of go through the process. But in terms of pre-training and post-training, you're always on the lookout for what gaps are there in the model, particularly to beat benchmarks, you know, that everybody is kind of competing for and looking on the lookout for all of that. So in terms of how does it work for us? Well, what really happens is that the labs need the data yesterday, almost every single time. So they expect turnaround times from project specifications to delivery anywhere from 24 hours to 48 hours to that time frame. So we have to be very, very agile to be able to deal with this.

Dhruv Sharma: And since the time that you've had a front seat view into all of this, how have you seen the challenges that, you know, the Frontier Labs, their state-of-the-art models face? Like how have the challenges evolved? I mean, you know, the models weren't good at something two years ago, good at that thing, but they're not good at something still. So how has all of that changed over time?

Sajo Mathews (ML Lead, Deccan AI): Yeah, I think, I mean, models have advanced in many ways, right? So if you even look at what our own work was.

Rohin Parkar (Co-Founder & CEO, Spintly): What are they still bad at?

Sajo Mathews (ML Lead, Deccan AI): Oh, so there is a lot of work which has to go today on making sure that models are able to autonomously operate in settings without constant, you know, human oversight, right? And that's what all the Frontier Labs are chasing, which is really, I mean, if you break it down, it is really about being able to do better tool calling, better contextual understanding and, you know, appropriate escalation. So that's where almost all models are trying to become better at today. But this has evolved from the past, right? Like when things started, it was just to make sure that the model output, which is the question that is being answered, is preferable to a human being, right? So just the preference was what things started with, but today it's evolved into trying to make the models run autonomously for long durations of time.

Utsav Somani: And I think the number that stood out in the introduction was that one million contributors on your network. How many of them are active every month? And what are the engineering problems or just challenges that you need to sort of solve for managing such an active contributor network? And only in one geography, I believe some of the other data labelers or training data providers like Scalia and Mercer work across different continents as well. So why focus on just India? Is it to reduce some of this complexity?

Sajo Mathews (ML Lead, Deccan AI): Yeah, so I think, I mean, I'm sure I think Ruklesh might be able to answer that better. But the context really for us is that, I mean, we are all about speed and quality, right? And for us to be able to deliver on the quality promise, I think what we really felt was that we could leverage the inherent scale that we have in India, where we have experts across so many different fields who are able to contribute. And yet we can have much stronger controls on how people operate to deliver on the quality front. So this is actually pretty unique to what we do, where we are a little bit more, we have a much larger base in a concentrated geography like India, which is mainly to improve our quality posture more than anything else.

Utsav Somani: And by quality you mean margin of error is close to zero or has to be zero?

Sajo Mathews (ML Lead, Deccan AI): Margin of error has to be zero for post-training data, right? So that's really what the enterprise is, I mean, for the frontier labs come to us for. And that is what we need.

Utsav Somani: For a second and welcome Rukesh as well to join us on the show. Rukesh, welcome. Thank you guys. We've gotten started without you, so feel free to chime in whenever you feel you have something to add. Good to have you with us. Yeah, thanks for having me.

Utsav Somani: Sorry, Sanju.

Sajo Mathews (ML Lead, Deccan AI): No, I was saying, yeah, so margin of error is pretty much zero. That's what the frontier labs need, particularly for post-training data, which is, you know, the point at which the models are extremely sensitive to quality. And yeah, our decision to kind of focus on one geography was really so that we can make sure that we deliver on the promise of speed and quality. Right. And that's that's really what, why it was picked in that sense. And Rukesh, you can add to that.

Rukesh Reddy (CEO, Deccan AI): Yeah, I mean, India has a seventh of the world's population and some of the smartest people around. So we didn't ever feel the need to go to 100 different geographies. I think just focusing on this one country and getting it right is far more valuable to us as well as our clients, the labs.

Dhruv Sharma: I'm sure like, I mean, a million human evaluators and they all have different specialties. But how do you guys harmonize the efforts of a million individuals who are all part of your network? Are there are the guidelines? Is there like, I mean, we'd rather have that answer coming from you.

Rukesh Reddy (CEO, Deccan AI): I mean, there's a lot of AI in there. We've got like AI screeners, AI interviewers, AI draw list, essentially the entire stack, which helps us find exceptional people. And then once they're in, we also kind of evaluate them using a ton of AI. And then there's also the human touch once they are closer to the process of actually working on live projects. So what we realized is we don't need to kind of coach people saying, hey, this is what a great prompt looks like, or this is how you evaluate a model, these are rubrics that you need to worry about and so on. And then in that process, right, we realize that all of us, right, including probably you, me, Raja, everyone on this call, we tend to zone out whenever there is something which is recorded and not live. The engagement levels for live sessions are always much higher. And so a lot of our L&D work, we actually deliberately do it high touch saying, hey, we've got a trainer who's working with you and telling you how to think about rubrics and whatnot. So there is that human touch that we sprinkle in as well, which is again one of our cool, part of our playbook has been that we are not this anonymized shop where you don't know who you're dealing with and we're just clicking away and working with a computer. We're like, hey, there is a lot of AI, but we do make it very real for you. You haven't asked me anything until Saturday. A lot of the leadership shows up. We do community bonding events and lots of other things to kind of bring together the best of AI and the human touch.

Utsav Somani: Rakesh, I want to ask you one thing, since you've joined slightly later in the conversation, I think your background is also fairly interesting. You were leading digital transformations at various banks, I think JP Morgan, Citi, and then in 2024, you decide to go into post-training data and start an AI company. What was the decision or what was the unique insight or what was the moment where you felt that this is an opportunity that's worth pursuing?

Rukesh Reddy (CEO, Deccan AI): Yeah, so I had always been trying to start up for, I don't know, probably a decade now and then considered many, many fintech ideas, which I had spent much earlier in finance. And then, you know, I never quite liked any of the of the ideas in fintech. And meanwhile, I almost stumbled my way into the AI and annotation and data category and it just blew my mind. I was like, wow, this is an idea ripe for building. And I mean, at the time, it wasn't as ML heavy as it is today. Like today, it's all about RL environments and AI stack and whatnot. Back at the time, it was also more about how do we find great people. So there was a lot of human ops, which still remains important, but it was maybe 80-90 percent human ops, 10-20 percent ML. Now it's 50-50. And probably by next year, it'll be like 80 percent. It's not next year. Even this year, maybe 80 percent RL and 20 percent human. But I wasn't shy, like off the notion that, hey, we've got to go heavy on human ops. Now it becomes fashionable to say, hey, services are software rather than software are services. But I think for me, being a non-traditional entrepreneur, labels didn't matter. And therefore, I could see the size of the opportunity and jump in.

Utsav Somani: And crazy growth as well. Congrats on the recent milestone with A91 leading their first AI investment into your company as well. And you've hit 10x year-on-year growth, 25 million in ERR now.

Rukesh Reddy (CEO, Deccan AI): Thank you. No, 25 million is the size of the fundraiser, we aren't disclosing the size of the revenue, but it's moving up every month. So yeah, hoping to go 10x and another 10x this year. So we'll be into the triple digits. So let's see, fingers and toes crossed.

Utsav Somani: And you're working mostly with Frontier Labs right now, or do you have other clients as well?

Rukesh Reddy (CEO, Deccan AI): We do have other clients, but a fairly large chunk of the work ends up being the Frontier Labs, because we are still in the, maybe towards the beginning or the end of the first era of AI, which is, hey, how do you help build these models? That'll continue for the next foreseeable future. But meanwhile, inference will continue to grow, which means how do you deploy these models into the real world settings? So that's where we are beginning to see a lot of enterprise interest. We have some really cool logos there as well. And the nature of work itself is also evolving, which is where Sajal is coming in. So it's not just about now creating the right data to get the models to be better. It's about saying, hey, how do I create a full stack agent solution, which is actually moving the needle for the enterprise?

Dhruv Sharma: How do you guys feel about synthetic data today? And also, what's it likely to be like, I mean, you know, two, three years into the future, what might it be like?

Rukesh Reddy (CEO, Deccan AI): Yeah, I mean, I think it's already here, the dawn of the human-AI collaboration. I mean, maybe two years back, we needed to create data which was entirely human because we couldn't quite trust synthetic data. And now we are at an era where pretty much most of our work ends up being, hey, let's get the AI to take a first pass. And then the human comes in and reshapes the trajectory as needed. So it's the best of both worlds. And even in our RL environment, like in Sajal, you can talk to this much more intelligently, but you're kind of training the environment using humans who are highly qualified and who have that taste or judgment or whatever you want to call it. And then the machine goes and does the rest. So this is just, we are already upon that future where it's both the human and the machine collaborating.

Sajo Mathews (ML Lead, Deccan AI): Yeah, I mean, just to add to that, right. So synthetic data really becomes, I mean, the value in synthetic data is only when you can take synthetic data, say, generated from, you know, some existing models or whatever it is, and then curate it to find the good examples, which becomes training data for the next iteration, right? So that's how synthetic data helps improve the models. We just take synthetic data without the judgment. It will just lead to model collapse. So judgment remains a very, very critical part of synthetic data. So you still need, you know, some mechanism, whether those are automated in some form where people are building verifiers or whether there's actually humans judging every single case, right? So even to derive value from synthetic data, I mean, you still need a lot of judgment. And that's kind of where we still do a lot of work.

Dhruv Sharma: Didn't Frontier Labs have like some bad experiences with, I don't know, data contamination and model collapse, as you said, too, for synthetic data in the early days and kind of still very early to get a bad name?

Sajo Mathews (ML Lead, Deccan AI): Yeah, I mean, if that's what's at stake, if you don't control what kind of synthetic data you use for model training, you could very easily kind of end up in, you know, a vicious cycle where the models just progressively learn all of the bad signals that you have in the original version of synthetic data, right? So the only way synthetic data can help improve a model is if you're actively distilling or actively filtering out the bad data out of the synthetic data and only feed the good data back into training loop, then it becomes a positive cycle, right? So so that judgment is like super critical, even for synthetic data.

Rukesh Reddy (CEO, Deccan AI): And to put it another way, right, let's say we've created a hundred trajectories, 90 of them end up not needing to be touched by a human at all. My human looked at it and said, wow, this is perfect. I couldn't have done this better. And the remaining time got touched. But guess what? The value of the entire hundred trajectories is like the human having to review those 90 and saying, hey, this is perfect. That's valuable, because if you were to take that entire hundred and say, hey, I'm going to directly go feed it into my training data set. And then, you know, the model starts misfiring. It doesn't help anyone. So so we are increasingly seeing that the value of a human is in bringing in that verification layer. And even it might be as simple as saying, you know, this is perfect. There's nothing wrong with it. That in itself is valuable. And we are still talking mostly frontier labs. The minute you talk about enterprise, the tolerance for failure is like 0.01 percent or maybe even less than that. So so, yeah, it will look like, hey, I don't need humans. But the reality is you do need the humans because you want to catch that edge case, which is going to otherwise tank your deployment.

Utsav Somani: So post-training data, I think, is a big enough use case, but are you exploring other product lines as well now that you have relationships with frontier labs and enterprises as well?

Rukesh Reddy (CEO, Deccan AI): Yeah, I think post-training data is very much in the realm of, hey, how do you help build models and get them to be better? Stage two is how do you deploy them? And that's where Saju and his colleagues are working out that enterprise angle. And by the way, by enterprise, we also mean like if you are a Google, right, you're building the Gemini set of models, but you're also going to deploy Gemini in your internal use cases. You've got HR ops, you've got ad operations and whatnot. Same thing for Meta, you've got content modification and whatnot. So so everyone's got to figure out how to use these models in an enterprise context. And and we've kind of already started on that path and we've got a bunch of exciting projects. So over to Saju to talk more about that.

Sajo Mathews (ML Lead, Deccan AI): Yeah, I mean, so on the on the deployment side, like I think one of the real challenges which people are facing today is it's you know, with all the advancements, it's super easy to build an agent. Right. You could maybe even do it over a weekend, but to be able to deploy it to run in a safe setting, right, where you know that you can trust the agent not to make decisions that would adversely affect your business reputation or maybe your, you know, regulatory compliance or any of that becomes extremely hard. Right. And that process involves people building all kinds of guardrails into the agentic setups to make sure that all these edge cases are handled. And that's where we see that there's a lot of work which we can done, which which can be done, particularly because we've seen what the models are capable of. We know where they fail, how they fail. Right. And we can use that learning to kind of build systems which address these failures at the systemic level rather than in a reactive loop. Right. And that's where we feel that we can now take all of this learning and deploy and help enterprises of all types deploy these models safely, basically shortening the gap from pilot to production. So that's that's really what we're trying to solve.

Rukesh Reddy (CEO, Deccan AI): I'll say it another way, like self-driving cars have been around for maybe 15 years now, and the first Waymo took the first unsupervised ride in SF probably 2012 or something. And even today, you've got like, what, 3% coverage in the US cities. And obviously in India, it's still much, much farther to go. And the reason for that is the cost of failure is catastrophic. It's almost fatal in the case of a selfdriving car. Right. And enterprise, the cost of failure is maybe not death, but it's it's pretty bad. Like you've sent out like 100K instead of 10K or you've deleted a code base, whatever. Right. So this is where I think it's super important to get that accuracy. Right. And beyond a point, it's almost like, yes, you can continue to build out accuracy to a very high degree. But then this is again where the human in the loop comes in to save you. And that's kind of the model we are building out, saying, hey, exceptionally accurate agents. But at the same time, we also bring in the human in the loop so that you can deploy in peace.

Dhruv Sharma: I think the self-driving car example is a great opportunity to ask the question of like the question, what about data for physical AI? Because I mean, data for generative AI, sure, you want an accurate outcome. It's still a probabilistic outcome. That's OK. But in physical AI, you can't have a probabilistic outcome that is not accurate. So are you guys going after that opportunity? Are the people going after that opportunity? What's the right way to think about it?

Utsav Somani: Also, I mean, I don't question to that. A fun one. Is Tesla the biggest collector of physical data set?

Rukesh Reddy (CEO, Deccan AI): That's an easy one. I thought you were going to ask Tesla versus Waymo. Yeah, I mean, I think. I don't know even what to call it, right? It's Tesla plus X plus XAI plus SpaceX. It's all going to be one big thing.

Utsav Somani: Yeah, probably the largest data set from space also, I think.

Rukesh Reddy (CEO, Deccan AI): You bet. No, but it's super exciting what's happening on the physical side of things. I've been following Yann LeCun for a few years now, and he's always said, hey, like language models, the natural evolution is that they need to understand the physical world. And that's when they'll get that common sense that we've only gained over billions of years of evolution. So, yeah, I mean, we are building out of physical AI practice. Hired our first few researchers. And meanwhile, we've got a tremendous amount of inbound interest and we've already done commercial projects. And yeah, I think there's tons to be done. Egocentric, egocentric, classical annotation, simulations. And we are kind of doing all of this. And still very much sunrise stage for this era. But yeah, very exciting to see how this entire space is evolving.

Utsav Somani: I mean, another fun question, actually, have you read the have you picked up a copy of the DeepMind book, which just came out from Sebastian Malavi, author of Demis Hassabis. I think that's a good one, because I remember through highlighting that documentary and I ended up watching it, I think, a few weeks back. And just the story of DeepMind, I think it's fascinating. And some of the origin story and how Elon Musk was involved as an investor and then eventually had to break away and wanted to buy it before Google did. And then he ended up starting OpenAI because of that reason. Just to not have to stay in the loose. So I think a very fascinating story. I mean, there's like some really Hollywood level stuff going on behind the scenes with this Frontier Labs.

Rukesh Reddy (CEO, Deccan AI): So that's going to be next on my flight reading.

Utsav Somani: Thanks for that. Sebastian Malavi, I don't know what the Infinite Machine or something it's called or this thing, but it's on the DeepMind founder, Demis Hassabis. OK. Yeah. Dhruv.

Dhruv Sharma: Yeah, guys, I mean, when you talk to your friends who are technical leaders in other companies, how do you guys chat about, you know, not slowing down and using everything that AI has to offer, but also not accruing technical debt in the process because some things still don't work?

Rukesh Reddy (CEO, Deccan AI): I'll take a quick first pass and then Saju, you can talk to this much more intelligent. I mean, Saju, I think has seen how do you build a startup, but also how do you build cool stuff inside an enterprise? So probably much more useful answer. But from my perspective, right? I mean, I mean, the Brave New World, I've been cloud filled. And I think the more people get there, the more they realize, hey, it's not necessarily a tradeoff anymore. Right. You can build and you can build guardrails and you can do it all yourself. You don't like you just instruct cloud overnight, wake up in the morning and you're whatever you're instructed to do. It's it's ready and you're just super powered, empowered. And yeah, so that's my take. But what do you Saju?

Sajo Mathews (ML Lead, Deccan AI): Yeah, I mean, so I would probably answer it in two levels. Right. One is just building things out using AI today. Yeah. I mean, you have you are trying this, but there is a real risk that if you kind of don't invest time in understanding what is being built, then you'll either have maintenance cost over the long run would go up drastically. I think you take a lot more effort to maintain a code base, which you have no idea how it got created than if you had actually spent time doing that review up front. So that's a real risk. And, you know, everybody kind of using AI to code has to be aware of that risk. But at the same time, I would really kind of equate this transition in some form. Right. When people started coding, like you had, you know, those punch cards and then people started doing assembly language and you would write those specific instructions, moving memory from one point to the other and so on. And in a way, you should really think about the new transition as saying that, hey, the current programming languages are one level of abstraction which exists to explain to the computer of how to solve a problem. And now we have moved to the next level of abstraction, which is saying that we are saying what we want to solve. And we're letting the LLMs basically act as the new version of compilers to compile that into source code and then, you know, everything else. But that means that we have to also build equivalence to, you know, test cases and, you know, whatever we will do to verify if the code base is actually correct in this new abstraction. Right. We probably still need to think about how do we specify our intent in the right way, which kind of LLMs can compile properly. So there are, again, a lot of work happening in terms of being specification driven and so on and so forth. So a lot of this is still evolving, but it's everything is still being figured out. Nothing is fully solved yet. But I would really say that that's the direction we are going, that we are actually building a next level of abstraction. So the problem still remains. We're just operating at a higher level now.

Utsav Somani: And I think a geography question as well. So what is India's place in this future that is AI built? Like, do we supply post-training data? Do we work on improving our foundational models? Where do you think this is two years down the line, three years down the line?

Rukesh Reddy (CEO, Deccan AI): I think the borders will only shrink. And I think there's always been this distinction between, hey, like, let's say the Western markets, the front office. And then a lot of what India has been doing is the back office or the middle office. And I think AI can do so much more, especially, I mean, with the current state of affairs, definitely back office tasks, middle office tasks get much easier. But then, you know, I mean, give it a few years and every task, right? Anything that we've done on the white-collar side of things becomes a fair game for AI. And I think there's two ways to look at it. Obviously, it's a risk to everyone. And, you know, if you look back at the history of mankind, it's a history of technology in essence. I mean, without tools, we were just monkeys, right? But so we've always kind of built our technology, which has replaced something which was manually done before. This time, it's scary, because it's for the first time, probably ever, you're trying to say, hey, white-collar work gets affected. But I think it's also a very cool opportunity, because you still have that judgment and that taste and whatnot, which you're able to bring in the systems. And I think India has exceptional talent, folks who understand, you know, how to operate inside an enterprise, how to kind of get the right stuff built out, how do you manage risk and so on. So ensuring that our glorious talent base is upskilling and working in this new world, I think that's going to be important. And yeah, I think we are, in our own little way, trying to do that, saying, hey, let's create the future of work. And let's create these new kinds of employment where you bring in your humanness. And then you're still collaborating with your AI partners. So that's, I think, a pretty cool opportunity. And I mean, yeah, I'll not talk about sovereign models and so on. I mean, that's a geopolitically loaded topic. And who knows, right? I thought the war will be over in three days. And here we are. So I don't know what's the right answer there. But I mean, if you take the other view, which is, hey, it's all one big, large global market, then I think there's an amazing opportunity for India to kind of help in deploying and so on. Anything you would add, Sagar?

Sajo Mathews (ML Lead, Deccan AI): No, I would just say that, you know, nobody really knows what the opportunities are. It looks scary, right? But unless we are in the game, we will not know what those opportunities are. So I think just being a part of the journey is super critical so that we can, you know, identify these opportunities and build on it and find our place in the ecosystem. So that's kind of where I would come from.

Utsav Somani: I think that's it from me. Dhruv, any final closing question?

Dhruv Sharma: I want to ask you guys, what's the Indian ecosystem or diaspora looking like in the Bay Area right now with everything that's going around? And how long have you guys been there? Have you been there for quite some time now?

Rukesh Reddy (CEO, Deccan AI): So, I mean, a lot of pride in, you know, building Deccan. I mean, the name itself. I mean, I've seen, like, Cisco is named after San Francisco and Menlo Ventures and whatnot. I was like, why is no one not naming a cutting edge company based on an Indian geographical landmark? So there you go.

Utsav Somani: Sorry, we'll keep the story behind the name before we start this whole thing.

Rukesh Reddy (CEO, Deccan AI): Okay, very cool, very cool. Yeah, so, I mean, yeah, I think the Bay Area is full of Indian diaspora and killing it, people out here. I've been in the US since 2014 now, and the first decade or so was in New York, which was my finance career. And now I split between the Bay Area and Hyderabad. I spend a reasonable amount of time in both locations. I mean, we are trying to be very close to the clients, but also be very close to delivery. And then, yeah, I think no better place than out here in the West Coast, where people recognize the value of smart people, irrespective of what's your heritage and so on. So as I said, I mean, I do think it's one global world where borders are shrinking.

Utsav Somani: All right. Thank you so much for joining us, guys. Wishing you all the best on this journey. Thank you. Thank you so much. All right, listeners, that's it from us today. There was not too much to cover in the news, but Apple has a new CEO, OpenAI has a new model out, and we'll see you on Friday at 4 o'clock. Thank you so much. Bye-bye.