Ashutosh Singh
RevRag.AI | OCTOBER 29
Builds multilingual AI agents that work inside banking, insurance and fintech apps, guiding customers through onboarding, answering product questions and assisting with sales and support.
transcript · reviewed NOVEMBER 18, 2026
#episode 156 transcript
RevRag.AI | OCTOBER 29
Builds multilingual AI agents that work inside banking, insurance and fintech apps, guiding customers through onboarding, answering product questions and assisting with sales and support.
Flam | OCTOBER 29
Builds AI models and products for interactive content, including 3D experiences and visual agents that people can interact with through touch and voice.
General Autonomy | OCTOBER 29
Robotics and AI company developing affordable robots for everyday tasks and AI foundation models for robotics that operate in the physical world.
General Autonomy | OCTOBER 29
Robotics and AI company developing affordable robots for everyday tasks and AI foundation models for robotics that operate in the physical world.
10,150 words
Dhruv Sharma: Today we're speaking with Farid and Bhanu who were previously founders of ShareChat and now they're building a company called General Autonomy. Guys, it's an absolute pleasure to have you on The Offline Network.
Farid Ahsan (General Autonomy): Thank you Dhruv and thank you so much for having us.
Dhruv Sharma: Bhanu, who's there on your screen?
Bhanu Pratap Singh (General Autonomy): That's our latest version of Humanoid. It's mostly an aluminum frame. Some parts are 3D printed. It has 23 degrees of freedom that is 23 actuators. Six in each leg, five in each arm, one at the torso and inside the torso you have battery and a single board computer. You also have some sensors around which is not there in this particular version but there are cameras which allow the robot to perceive the area around it and do complex terrain navigation etc.
Dhruv Sharma: That's a lot of engineering but maybe we start by asking... Oh man, it's moving, it's moving, it's moving. Oh, it's moving.
Utsav Somani: It's doing a full... Oh, it's walking away. Coffee break time.
Bhanu Pratap Singh (General Autonomy): As you can see, the walk is pretty human-like because we train on human references. We put in available human videos, extract references and then put that in simulation and then train and reinforcement learning on top of it and the result is this.
Utsav Somani: Guys, a very sidetracked question. There was this race of humanoids, I think somewhere in China or maybe Korea. Do you think you're naming this one Tara? So will Tara participate in that soon?
Farid Ahsan (General Autonomy): Well, if the Chinese government allows us to bring Tara without revealing our IP, then why not?
Utsav Somani: Alright, let's get into it. Dhruv, over to you.
Dhruv Sharma: Yeah, I mean, here's the question I've been waiting to ask you guys. Why robotics? We'll get into the engineering of all of this again but after the success you've had with ShareChat, why robotics?
Farid Ahsan (General Autonomy): Yeah, so Bhanu and I, we used to build robots back in college. Like the reason we became very good friends, you know, was because we loved building stuff and we loved building robots, we loved building systems, we loved building apps. So this, the love of building, you know, got us to bonded, you know, over many projects we did and when we were graduating, we actually did want to build a robotic startup together but 2014 was not the best time to build a robotic startup. You know, building a robot, a very simple robot, that time would take you about three months. Compared to that, doing that today is probably like a few hours job, right? So we believe that there would be a robotic future someday and we just wanted to be ready with, you know, the right set of skills, the right set of, you know, business mindset, technical mindset to be able to take the opportunity with both hands and early 2023, both of us figured that robotics is at an inflection point where the chargeability moment for robotics would happen within the next three to four years. So, you know, for us it just made sense that we should jump in into robotics again that time, yeah.
Bhanu Pratap Singh (General Autonomy): Also in 2023, a lot of new research papers came out that made it easier to do robotics. There was this one paper on manipulation, which uses a diffusion policy. Then there were a bunch of papers on locomotion, using reinforcement learning. They made it easy to do robotics. Before that, robotics required a lot of different layers of systems, you know, coming together to solve the task, whereas now robotics is about creating one neural net at the end, and just running that neural net at 50 hertz.
Utsav Somani: I mean, I'll let all the technical stuff be covered by Dhruv, but I think the supply chain is also a very interesting part. You publicly stated that most of this is actually made in India, apart from like NVIDIA Jetson and maybe one more component, I think, one of the motors or maybe some parts of the motors I think are imported, but everything else is Indian. So how did, how tough was it doing all of this in India?
Farid Ahsan (General Autonomy): The problem is not that, other than obviously the NVIDIA compute and the Intel compute, the problem is not with, you know, manufacturing any of it in India. The problem is that India does not have a thriving prototyping ecosystem. So if you want to experiment, it becomes very hard to find vendors who are willing to provide you very low MOQs, to test out things, and, you know, get your experiments verified, or get up to speed with the global standard very quickly. So that has been one of the challenges. Our motors, our actuators are entirely made in India. We do have bearings, cross-roller bearings that are imported because very few people in India make cross-roller bearings of the size that, you know, we use in our robots. And the magnets, the permanent magnets that go into the motors, we do not have any indigenous supply of that. So we have to depend on either China, Taiwan, or Korea for the rare earth magnets, that's the permanent magnets. It's actually an alloy called NdFeB, neodymium iron boron. So yeah, other than those, everything else, we were able to make them in India.
Utsav Somani: China has that rare earth export ban or pause right now, right?
Farid Ahsan (General Autonomy): Yeah, so that's true. But, you know, we've been able to figure out ways with which we can, you know, get some supply. And then there is a lot of growing, I would say, industry output around secondary supply chain, where, you know, all these rare earth metals are being extracted by the likes of companies like Luham and all who are, you know, basically recycling smartphones, recycling laptops, and, you know, extracting these rare earth metals. So there are ways with which we can serve our needs. And, you know, currently, we are not producing at scale. So for our prototyping needs, we are well taken care of. But I agree, as we go forward, this will be one of the big challenges that we'll be, we'll have to solve.
Dhruv Sharma: Fantastic. I think, guys, to understand this stuff a little bit better, what would be super helpful is if you could explain to us like the hardware subsystems, and the software stack both of the quadruped as well as the humanoid we have there on the screen. Sure.
Farid Ahsan (General Autonomy): So I think the biggest difference like I let me start from the difference between the two robots, the biggest difference between the two robots is the number of degrees of freedom per limb. So a robot dog normally has three degrees of freedom per limb. So that gives it a 12 degree of freedom, you know, for your system. A humanoid, on the other hand, has at least five degrees of freedom per limb and can extend to six. Like in the case of our humanoid here, you've got six degrees of freedom in the legs, and five degrees of freedom in the hands. And that, that exactly dictates the kind of applications that they're useful for, right. So when you want to do many complex manipulation tasks, like, you know, when I say complex, it can be anything from cooking food to, you know, bringing objects, playing football, playing cricket, playing tennis, helping somebody, you know, follow their fitness regime, or basically be a fitness coach or, you know, anything under the sun that you can think of. But when you reduce the number of degrees of freedom, you basically reduce the number of applications. So the dog is capable of locomoting to any place, you know, while, you know, while walking, it can, it can go to any place, it can get into any orientation, but that's about it. That's all that it can do. Right. So subsystem wise, I would say, there's the, there's the single board computer, the electronics. So we've designed our own single board computers, there's the battery and the battery management systems, there is the motor controllers, which are attached to each of the actuators, there's the actuator itself. And then there's the structural parts. Other than this, you have a few sensors, like cameras, IMUs, in some cases, depending on the application, you might want to put a LIDAR, but we've been able to overcome that requirement. Now, I'll let Bhanu brief on the software requirements.
Bhanu Pratap Singh (General Autonomy): I'm more hardware. So there are essentially three components in the software. One is the sensor stack that contains your camera, IMU, LIDAR, etc, all the sensors. There's some pre processing in there just to remove noise and you know, model the uncertainty so that we can fuse different sensors together. This is the sensor stack. Then when it comes to the brain, there are two components. One is a high level network and another is a low level network. High level network is more about if the robot wants to do a certain task, how does it break down into subtasks? And whether the robot is able to do those tasks? Are there essential ingredients available around in order to do that task? All the reasoning involved goes into the high level network. The low level network is more about the execution, where the network generates the actions for the robot to accomplish a certain task, their precision, failure, retriability, etc, are the main components. Both of these are neural networks, which are trained in Muslim simulation, but there are real world datasets as well that you can use in the offline training stage. All of this then comes together as the final brain of the robot.
Utsav Somani: And what are actuators? You mentioned motors and actuators also designed by you in India. So maybe for our audience and even me, what are actuators?
Farid Ahsan (General Autonomy): So an actuator is, it's a joint, it's a joint module, where if you look at the robot closely, Bhanu, can we bring Tara closer? Yeah. Yeah. Yeah, if you bring closer and you go closer, if you look closely towards the shoulder.
Utsav Somani: Do you give voice commands to Tara or is somebody controlling it right now?
Farid Ahsan (General Autonomy): No, right now it's being, it's being controlled through a remote, we've made a remote control for this. So anyways, if you look at the shoulder. Can you make Tara wave to us and our audiences? You should have told us before.
Utsav Somani: Next time you hear, then we'll try some tricks. Yes, yes.
Farid Ahsan (General Autonomy): Actually Tara right now is more like Thakur from Sholay, he doesn't have hands. So anyway, the actuator is essentially a motor with a gearbox, with a encoder and a motor controller. And you can put all of this in a shell together. So an encoder basically tells that once you have given a signal to a motor to move, let's say 30 degrees, if it has actually moved 30 degrees or it has moved 30.1 degrees or 29.9 degrees. So this verification of how much the motor has rotated, this is done by encoders. The motor controller is the one that takes in the, you know, the system command and converts it into this action of rotating. The gearbox that is attached with the motor allows that rotation motion to be converted into the necessary torque required for lifting the body. And within the motor, you can actually still split it into two parts, the stator and the rotor. Now, the design that we have taken requires one of these two to rotate and the other one stays static. Different people use different, you know, styles of stator-rotor design. The one that we use, its technical term is permanent magnet synchronous motor. And it's a, you know, niche version of a brushless DC motor, which I'm sure you must be aware of.
Dhruv Sharma: Yeah, I think so. I mean, super interesting so far. Farid, earlier you'd spoken about the robotics supply chain in India not being, I mean, not being there yet. How far has it come already? I'm sure even in the three years you guys have been active, you've seen some signs of improvement. Tell us more about that.
Farid Ahsan (General Autonomy): Yeah, definitely. I think we are in a place where the stator, the rotor, the encoder, the motor controller, this is entirely solved in India, except for the magnets. So leaving the magnets, everything else, you can get it in India. The gearboxes, where a lot of precision engineering goes into, it's not that we do not have the supply chain in India. It's just that our current supply chains are very much overwhelmed because of huge aerospace demand in manufacturing in general. So if you look at any manufacturing company right now who's doing contract management, like right from the likes of the big guys like Zetwerk to, you know, new age startups like J Aerospace, their heads are completely filled with, you know, orders. So I would say that we are in a bad situation as an ecosystem. But we are early, but we do have all the ingredients. We have the science. Now, if India has to go through a robotics revolution, is it possible? Yes. Can a new team, you know, get up to speed with the latest in robotics and, you know, start matching, let's say, the Unitrees and the Teslas and the figures of the world? Definitely. It's probably like a year or two away from a point where you will see, you know, a Cambrian explosion of robotics all over India.
Utsav Somani: Nice. And I mean, give us the applications. Like you've developed these two versions, Tara, which was known as Atom 01 before and Param, which is the four legged robot dog. So where do you think they will be most useful in the Indian context?
Farid Ahsan (General Autonomy): Yeah. So I think while we have been in agrarian economy for a while and then we became a service sector based economy for a while, we are now seeing the signs of us becoming a manufacturing powerhouse as well. But the biggest challenge to, you know, this reindustrialization of India and, you know, becoming a manufacturing powerhouse is the availability of skilled labour, right? And skilled labour takes its own time to become very good and very adept with whatever tasks that they are given. The addition of robotics really allows skilled labour to increase its per capita throughput very quickly. So we do not see robotics as a replacement of humans because humans are far more creative, far more capable, far more able to take risks in situations which are which a robot or a LLM cannot really decipher. Like they will not be able to take those extreme calls yet. Right. And the availability of robotics allows a skilled labour to basically perform 10, let's say, for example, soldering tasks or, you know, 10 driving tasks like a screwdriving task where you are basically taking a huge component. You cannot use your existing smart tools that have existed in automated assembly lines and you need somebody to actually go and, you know, screw a big, you know, big bolt or something. Right. So a skilled labour will be now able to control 10 of these robots to achieve it on 10 pieces together so that they can monitor and make sure that all of those 10 are doing the job correctly. If we are not heading towards a complete lights out automation anywhere in the world with robotics, like I'm sure you must have come across examples from self-driving, you know, right now SF has a lot of self-driving, like this is just an example, self-driving cars available, but it's not that they are 100% self-driving. Those robotaxis have, you know, one driver sitting somewhere, you know, looking at 10 screens together. If all the robotaxis are driving well and if there's an extraordinary scenario that pops up, you know, the tele-operating guy just takes over. So now the earlier the driver was probably doing like just one robotaxi driving when he was, one taxi driving when he was actually a driver. But now as a tele-operator of robotaxis doing 10 drives. So we expect something similar for.
Utsav Somani: And this was an off figure as well, that robot, I mean, the humanoid that was demoed where somebody said, hey, somebody is peeking into my home because they have to control this humanoid to fold my clothes and other things as well. So it was a data gathering plus pilot in motion also.
Farid Ahsan (General Autonomy): Yes, yes. So we think that for humanoids to be able to get into houses, we need a lot more policy work, regulation work, insurance work to come in before we can actually expect, you know, a massive commercialization and adoption of robots in-house. Because you do not want a Saas, Bahu and robots TV series, right? You do not want that. You do not want it to be, you know, a reason for fights. You want it to be a reason for peace. So I think we are far from that point, but we'll get there as a complete ecosystem globally. I think probably two, three years at best. And we'll get there where we can leave robots in your house and expect them to perform well without any human intervention.
Bhanu Pratap Singh (General Autonomy): There are plenty of outdoor tasks that can be done, which have much lesser safety requirements. Firefighting or, you know, just doing logistics from one place to another, etc. These can be done. And even if the Indian market is slow to adopt humanoids, there are plenty of global markets in which these humanoids can be sold. And we can definitely create a supply chain in India and provide for all those global requirements right from today, rather than, you know, just waiting for this whole wave to be over and then eventually deciding that we want those supply chain jobs back in India. Instead of that, we can just pre-plan, put a lot of capex in and make sure that all the global markets are being served by Indian humanoids.
Farid Ahsan (General Autonomy): Actually, some of these decisions that we're taking is actually coming from our past experience. You know, when we started ShareChat long ago, FB was celebrating its 10th anniversary and Binance was probably celebrating its 7th anniversary. So for us to catch up to the global giants was taking a lot of time. This time around, when we started up, we wanted to make sure that we are not way behind the global state of the art. And as you can see, we have been able to match most of what the world has achieved at state of the art already.
Dhruv Sharma: Bhanu, my next question is for you again. On the software side, we'd love to know what experiments have you guys been doing with algos and with models to get the locomotion right, to get the navigation right, to get the manipulation of these robots right?
Bhanu Pratap Singh (General Autonomy): So we have been building this model called Latent Behavior Models. These are different than VLAs. These models allow you to train on a very large set of behaviors as a pre-training architecture. This then gives you a model which has at runtime all the skills that you need. And you just need a planner which will schedule those different skills in order to accomplish the task. Before this, VLAs and et cetera, all the slightly older architectures did not have this runtime configuration and did not have this vast skill library already available inside the model. All those training architectures were specific to a certain task. You decide the task, you collect the data and then you train it. Then there's also this issue with multitasking, getting all of that knowledge inside a single model is difficult. But behavior models start with that premise that they train on a very large data set of skills and those skills are available. Those skills are also precise, they are accurate and they are retrievable. Then it just becomes a job of a planner to schedule those skills to solve tasks. So we believe that going forward, we will mature this behavior model and then attach it to a high level planner. This high level planner can be anything. It can be like a GPT-Astra model or something that Claude comes up with. So this also offers pluggability with another reasoning models that are out there instead of us designing the entire stack from scratch.
Utsav Somani: I mean, there's a lot of debate happening about deep tech as well. So founders are, of course, at the forefront of doing stuff which combines hardware and software. So how much of the time within your company and within the product that you're building also goes to hardware, to software, to making a beautiful symphony between the two?
Farid Ahsan (General Autonomy): So, you know, Bhanu and I, we split hardware and software between us and that was how we used to make robots also back in college. So for us, it is a symphony we've been dancing to for the last 17 years. Right. So, you know, and the entire team is just an extension of the two of us. Right. So for us, it's, you know, till the time we two are in sync, the company stays in sync. And, you know, so far we haven't seen a massive hiccup there.
Utsav Somani: Was there ever a problem that you, I mean, you created something, CAD drawing or something and Bhanu was like, no, man, software can't for this. I mean, it's not feasible.
Bhanu Pratap Singh (General Autonomy): So our design process is very application. You can say it iterates in a way that before even machining anything, we have the results ready just through CAD. So you make one basic CAD design, then we put it into simulation, we test its capabilities. Then we also test what is what we can achieve using this model. And then that goes back into the design process. There is a further iteration. It is further again, checked with the software. And this iterative approach helps us, you know, avoid a lot of rework later on.
Utsav Somani: Do we have any final quick closing one?
Dhruv Sharma: I have a fun question to ask the two of them, which is, you know, to capture like the popular imagination of people, right? I'm not talking about sci-fi people, robotics, nerds, etc. Like Boston Dynamics, Unitree, they've always been very gimmicky. The somersaults, the backflips, the martial arts, etc. Is that the only way to capture people's imagination with robotics? Or do you think there's a different way to do it as well?
Bhanu Pratap Singh (General Autonomy): I think, yeah, Bhanu, go ahead. I mean, I would say they did all that because the model architecture was not there yet in order to solve actual problems. So those gimmicks definitely help. But at some point you need to solve actual problems and we are focused on solving actual problems. That way you won't see us, you know, performing these gimmicks for PR value, etc. You would rather focus on our customer or whatever the customer needs. We would rather want to do that compared to just a PR value stuff.
Farid Ahsan (General Autonomy): And just to add to that philosophically, I think when people look at robots, they want it to do superhuman stuff. So a backflip is like a semi-pro superhuman already. So I think it just helps people understand that, OK, this can actually supersede a human when you just show them doing, you know, regular bartending or regular pick and place. It doesn't excite the imagination because you don't really imagine, OK, this robot will not take leaves or won't get sick or, you know, won't work in shifts. It will just keep working. So it doesn't fascinate enough people. But backflips do that. Frontflips do that.
Utsav Somani: Guys, this was super fun. Thank you so much for coming on our show. Wishing you the very best ahead on this exciting journey.
Farid Ahsan (General Autonomy): Thank you. Bye Tara.
Utsav Somani: Oh, Tara's walking away. See you, Tara. All right, listeners, moving on to our next guest, we've got Shourya from Flam. Shourya, welcome to TON. Hey, Shourya. Are you dialing in from SF? It seems dark.
Shourya Agarwal (Flam): Yes, I'm dialing in from SF.
Utsav Somani: Up early for us. So we'll keep this fun. I think the fun aspect is I went to your website. I see that you raised 40 million from QED and it also mentions and Shah Rukh Khan. So was Shah Rukh Khan a strategic investor? Like what's the benefit that you're getting out of him on your cap table?
Shourya Agarwal (Flam): I think the king of content should be investing and, you know, be closely associated with the next big thing in content. I think that was the intent. Yeah. And I think we have a significant business in India. So we just wanted to have more brand associated to Flam. But yeah, I think the main reason is like this is a new modality of content.
Utsav Somani: And I think he's like the best in that space. Perfect. But can you describe for our listeners who are hearing about Flam for the first time, some of them might be running businesses also or content companies or large, I mean, working at large companies. So what is Flam and why should they be interested in it?
Shourya Agarwal (Flam): Absolutely. No, I think it's a very exciting space. Flam essentially is a frontier AI research and product company defining the next modality of content beyond text, images, and videos. So internet essentially is nothing but like function of content. Everything is content on internet, whether it's like social media, entertainment, commerce, gaming, everything is content. And if you think about content, we've just seen like three modalities on internet, which is text, images, and videos. And if you see how internet has transitioned in the last 25, 30 years, we are basically moving towards more engaging modality. Like 25 years ago, everything on internet was text heavy or could messenger or code, yahoo, messenger, Facebook, and early days, all the websites used to be like text heavy. From text heavy internet, we transitioned into images on Instagram, and then we further transitioned into videos on TikTok, Dubslash, basically. So if you see this transition of content from text to images to videos, we are basically moving towards more engaging form factors. And we strongly believed a few years ago that this trend will continue and the world will move beyond videos into a brand new modality, which will be more engaging than a video format. That's what Flam is.
Dhruv Sharma: And sure, interactive content delivered to people's phones, devices, or wherever they're at, what would impact the foresee on marketing on advertising at a high level?
Shourya Agarwal (Flam): Yeah. So I think all the formats of the past are passive. Like you can't talk to a video, you can't talk to an image, or you can't interact with a paragraph of text. But if you could really engage with the content real time, and if you could navigate content, like we navigate real life conversations, that's like super engaging compared to anything that you're just like watching passively. I think it can be a video game, pretty much. Is that what you're describing? That's a very close way to look at it. And yeah, gaming is one kind of thing which is interactive real time. So if entire content could be like that, you will have like brand storytelling in a very different way. You will have like movies in a very different way. Commerce will change completely. So I think almost everything that we do with images and videos today can become like far more exciting if it's a flat format.
Utsav Somani: I think before we go into the various products and stuff, do you want to screen share and maybe show us like a demo or anything that you have handy? Like, because we're talking about an interactive content piece and interactive content companies, so it might be actually lovely for us to see something.
Shourya Agarwal (Flam): Yeah. Let me, let me share my screen to show you something. Um, that's cool. This is one of our, we have three different products, which is like Airboards, uh, Flix and visual agents. I'll just share my screen to show you a glimpse of like, let's say Flix. Flix are basically, uh, I'm just sharing my screen as I talk.
Utsav Somani: Uh, what brands are you live with currently?
Shourya Agarwal (Flam): Like, um, many different companies like Google, Samsung, Hyundai, many companies are leveraging Flam for different, uh, different things. Like if you see my screen, this is one of our products, which is called Flix. Flix are basically zero latency, interactive videos. So if I tap on any of this, uh, there's literally zero latency at which I can real time navigate a video content.
Utsav Somani: Uh, if you think about traditional formats, so let's say, uh, this is, this is all rendered at the back earlier, which will require insane compute power. Because if you have to do like, I don't know, a million combinations, then this thing.
Shourya Agarwal (Flam): Exactly, exactly. So basically if you want a more engaging format compared to a video beyond interactions, you have to solve for space, time, complexity. You basically solve space, time, complexity, either by doing TV content, which we do on airboats, which is like high fidelity 3d content, or you, or you solve for time complexity, which is like literally bringing down the latency to zero milliseconds. So if you see content on the left, uh, this is like the traditional HTML format. So if I try to tap on any variant, it takes like two 70 milliseconds to switch from one variant to another. On the right hand side is the Flam format, which I switch in like less than five milliseconds, which is like almost a hundred X faster compared to how actions could have been in the past.
Utsav Somani: Wow. And maybe when, et cetera, I think the latency even goes up. So maybe like a customer can drop off potentially, but do you have like real life case studies as well? Where like, I don't know, somebody has said that, Hey, our customer purchases have gone up because of this.
Shourya Agarwal (Flam): Yeah. And that goes significantly up. So Flam format has three X more watch time compared to YouTube ads, which is like massive uplift, 70% more recall. Obviously people are spending more time and interacting with it and not just passively watching. There's like much higher recall compared to anything we have seen in the past. And all of this leads to like massive ROAS, like ROAS is like the main parameter brands care about. So obviously people are spending more time and there's more recall. There's like much higher ROAS. So there's like a high superhuman performance, which we are like seeing across any kind of Flam format.
Dhruv Sharma: Sure. What customer insights does this feed back to the brands? You're talking about brands as a use case, like marketing as a use case? Yeah. I mean, for instance, you showed us that the, I think it was Zara and, you know, customer interacting with in that example, would they know that this particular individual likes red more than the other two colors, things like that?
Shourya Agarwal (Flam): Which is the most important thing for brands beyond just like the first time, a high increase in performance, like the kind of data, data that you can capture with like user interaction is like something brands have never been able to do before because when you passively consume content, you never know, like, what do I really want? But the moment you have consideration data, the kind of retargeting that that brands can do is like never seen before, because now I know that people in the Midwest, or let's say this particular individual likes red more than green, so I can show shades of red the next time when I'm showing content to the same individual.
Dhruv Sharma: Then it's almost like a flywheel, right? Like better interactive content means more engagement means more this sort of behavioral data means better retargeting and therefore better personalization.
Shourya Agarwal (Flam): But 200% it just like keeps on becoming better and better. Even the first time the kind of content engagement we are seeing is like much higher compared to videos, but like the data that you're able to capture, that changes the whole game on like how brands can further interact with the customers.
Utsav Somani: And you've gone through your fair iterations of different kinds of content, right? Smart photos, a metaverse, enterprise mixed reality, and now interactive content powered by AI. So what were the learnings across all of these evolution journeys?
Shourya Agarwal (Flam): That's a very interesting question, actually, because we did spend a lot of time, like almost like two and a half, three years researching and experimenting like what will the new modality beyond videos look like? If you think about it, it's a very hard problem, because like, it's almost very rare that an individual or a company will be able to introduce a new modality on internet or a new communication format on internet. So we did go through a lot of iterations, like a lot of experimentation, but I think finally in 2024, when we actually had our first product launch, we got to like 100 million views on the first plan. We still compare our first plan to the first video, which was launched back in 1888. So, so we feel that if we were the ones who innovated video as a format, we would be dead by now, but videos continue to become bigger and bigger in size and use cases and adoption. So if 2024 was the first year when the Flam format went live, it can, it can become like a much bigger thing and can outlive us. So I think when we saw that, we saw that in like, in a very big way, but until then it was like a bunch of experiments and research going on.
Dhruv Sharma: And I believe we have no less than five models that deliver this interactive content. Can you tell us more about them?
Shourya Agarwal (Flam): So basically new modality would require new, new models to generate that content. There are two sides of any new content. Like, let's say, when you talk about videos, you need a streaming infrastructure to stream content on a hundred percent devices. And then you need like generation models to generate that content. Similarly, when you, when you have like a new type of content, you need like new models to generate that content. So Airboards, for example, is like high fidelity 3d content, which requires a fable model to generate that content. Flix has a different kind of model. And then we have visualizations.
Utsav Somani: This fable is not the anthropic fable, right?
Shourya Agarwal (Flam): No, this is not that fable. We did think that we should like change the name, but we thought it doesn't matter, like the names keep coming. But yeah, fable essentially is like a 3d generative model, which is like high fidelity 3d content. And then we have like more models for different kinds of like products in the suite. And you can see the entire, like, like this is something, because you asked for like screen share. Like, this is one interesting thing that you should see.
Utsav Somani: You can have like a little screen share and show us all of these things. These are pretty cool things.
Shourya Agarwal (Flam): Yeah. You can basically have like different characters in the movie. You can have like two main characters. You can change the characters mid movie. You can insert yourself in there.
Utsav Somani: You think like interactive movies on OTT will become a thing? Like I remember on Netflix, there was one experiment where you could choose the direction where the movie is going, but you can't change characters.
Shourya Agarwal (Flam): Yeah. Yeah. So that movie was like Bandersnatch. Where you could like change the story midway. But the whole point is how, how much is the latency and like how real is the whole switch? So I think that changes the game on the user side, which is why zero latency or like sub five millisecond of latency plays a big, big game.
Utsav Somani: So latency is a key part of all of this. Right. And I'm guessing it's going to continue improving. So, I mean, when you start working within the future, Hollywood studios are large ones, will you have to deploy your models on prem? Is that something that you consider or it's all going to be delivered remotely cloud based?
Shourya Agarwal (Flam): No, we will have to like, we are considering on prem as well. Uh, on the, on the, that's on the generation bit, like when we talk about models, we talk about whether it's our, our server or on prem. But like streaming can be again, uh, both like on our server or like on the customer server. So all of that can be customized.
Dhruv Sharma: Do you think that changes for, uh, individuals using sort of low spec devices, Shourya?
Shourya Agarwal (Flam): So, so which is again, a very critical thing. Uh, it has to play on a hundred percent devices. So all our products can be streamed on a hundred percent devices. Uh, and I think that's the main requirement for any modality because you can't just like play on high end devices, uh, which is also something that took a lot of time. Uh, so in fact, it'll be fun to know that the first. Original coder for VLC media player. Uh, he wrote the code for video decoder and video streaming back in 1997. I'm sure everybody of us used VLC media player.
Dhruv Sharma: Yeah. Let's keep volume. As a people are like, Oh, 100%.
Shourya Agarwal (Flam): Like it could go up to 200%, I think. Absolutely. Yeah. And, and this guy essentially is running data dog, uh, uh, today, which is like a hundred billion dollar company. Uh, but I got to know that he was the original like coder on VLC media player. So when we introduced our streaming tech, like essentially Flam is a streaming technology. And then we ended up building up all the models, uh, in our stack. So we wrote to him and now he's also part of Flam very closely involved. Uh, so it's fun to see like after 30 years, a new decoder or a new streaming tech built for a new kind of modality.
Dhruv Sharma: I almost got carried away for a moment. Sure. I thought when you asked us, we know the VLC guy, you're going to say he's our CTO, but man, all of this must involve like serious, like tech work, right?
Utsav Somani: So patents and all of that stuff, you're being protected across the world.
Shourya Agarwal (Flam): Yes. So we, we have like patents and multiple geographies at this point. Uh, we first acquired like a bunch of like patents on the streaming side, streaming and decoding. So there are, as I mentioned earlier, there are three products on the suite. So all three products have like multiple nuances where you have to like, um, uh, think about IP, uh, and, and yeah, um, now we are doing the same on the model stack as well.
Utsav Somani: Will you supply this tech to Netflix or will you become a production or production house yourself?
Shourya Agarwal (Flam): So it's very exciting to see like multiple kinds of companies are now getting built on the Flam infra. Uh, I talked about branch storytelling and marketing as use case, but there's a whole social network getting built. There is a micro drama app getting built on, on Flam. So I think we don't intend to be the application layer, uh, essentially like the consumer app, uh, so to say. Uh, but yeah, like, uh, I think we keep providing the infra to all these companies.
Dhruv Sharma: Sure. You think for interactive to become the new modality, I mean, you guys are serving one third of the fortune 500 anyway, but you'd have to go to real taste makers, that's a new word we learned on the show, by the way, would, would they have to set the bar for this? Like, I don't know, Apple or, or others.
Shourya Agarwal (Flam): Yeah. I think, uh, enterprises, uh, are using it both for functional use cases as, as well as for fun use cases. Uh, Hollywood, Hollywood will use it for like fun use cases. Then social labs will be built on it for like fun use cases. I think all, almost everything that can be done with the video, uh, will be done on a Flam format. So yeah, it'll be both fun and functional.
Dhruv Sharma: I'm just saying that like Salesforce is not going to make this thing cool, right? Nothing against Salesforce, but you would have to get the cool guys to start using this also, uh, which I'm sure they are already. Yeah.
Shourya Agarwal (Flam): Yeah. I know we are getting like very cool companies to now use it. You will see more cool Flams coming out. It'll be boring. I can assure you.
Dhruv Sharma: What you showed us was far from boring.
Shourya Agarwal (Flam): Yeah.
Utsav Somani: Yes. What's it back to you? And, uh, how do people pay for you? Pay per view, pay per experience. Like you mentioned three product lines. So how do they pay you? Is it different across these product lines?
Shourya Agarwal (Flam): So Flam gets paid for streaming as well as for creation. So when you pay for streaming, you pay on a per minute or like on a, uh, yeah, on a per minute or a CPM basis. And then you pay for tokens when you leverage the models to create content. So it's both for creation and streaming.
Utsav Somani: Are there any final closing one?
Dhruv Sharma: I, um, I want to know more about the team as well. Sorry. And we asked this, if all our, um, you know, guests who are tuning in from that different from that other time zone, what's it like over there right now? Is, uh, AI psychosis peaking? Is it how often are people doing?
Utsav Somani: Well, I like robots.
Shourya Agarwal (Flam): Well, probably got doing really well, but and so it was like, uh, finally, like one minute, uh, walking from us. And we just like, you keep bumping into some of the amazing people from Anthropica and all, all these amazing companies, but honestly, like what's happening in SF is very unreal. I think at this point, uh, it's, it's borderline scary. Like literally every day, something changes dramatically. Uh, so you have to be like on top of it almost all the time. So me, uh, being awake at four is like, is a very usual thing.
Utsav Somani: No man. Sorry to wake you up so early, but, uh, I mean, QED investors, like that's a FinTech fund. Why is it investing in a media or a content company?
Shourya Agarwal (Flam): I think at this point, almost every fund is becoming an AI fund and everybody wants to invest in, uh, frontier AI companies. So I think, I think it's not just QED, but almost every fund you see, they are investing a big time in AI.
Utsav Somani: Alrighty. I think that's a good note to end the show on. Thank you so much. I have a wonderful day ahead.
Shourya Agarwal (Flam): Thanks, Utsav. Thank you.
Utsav Somani: Ashutosh, uh, welcome to TON. Hi, Ashutosh, you're our final guest today. Now we've had an AI special, so we're like, I mean, AI build at this point of time, we've seen robots, we've seen AI change content on the fly. So what are you teaching us today with RevRag.AI?
Ashutosh Prakash Singh (RevRag.AI): Yeah. So, uh, I think what we do at RevRag.AI is majorly building app, uh, app agents, or basically we, we, we have a thesis that apps are changing, the digital surfaces are changing, right. They are becoming agent tech. So what we do at RevRag.AI is that, uh, any app or any web surface that is there, we can convert it into an agent and with RSDK, with our, uh, uh, app agents.
Utsav Somani: Is this agent definition the same as like, say a muse agent or an instinct agent or like a dots or a cloud do it or something, or is it like you're using agent in a different meaning here?
Ashutosh Prakash Singh (RevRag.AI): Yeah. Right. So, uh, a good question, actually. So what we do is like, let's say there is an app of, uh, let's say Bajaj finance. Right. Now a customer comes in who wants to apply to a loan. Now, what he can do is he can just talk to that app and can complete the whole journey and get the loan done. Right. So, and even let's say he opens up that app and wants to understand or buy that burger of sort, right. He can just talk to that app and it will be, uh, doing the stuff and it is all conversational. It is all action intelligence led. And at the same time, it can like the agent inside the app can take action on behalf of the user navigating, guiding, and, and even showing, uh, uh, generative UI elements as well.
Dhruv Sharma: Ashutosh would, would this be a good way to understand this? Like there's many apps out there. Not everyone can afford to do like a AI native redesign. And plus it uses already used to the core product experience. And that is sort of layer one L one. And so you guys are sort of superimposing a layer, a second layer, which is, which is what you're calling an agent. Earlier people would call it a chatbot or whatever that uses the app on their behalf, behalf in an AI native way. Absolutely.
Ashutosh Prakash Singh (RevRag.AI): I think you're right. Uh, so we, we have two separate agents that we are experimenting currently. Uh, one is the co-pilot agent, which is basically if there is an existing app, let's say an HDFC app or an ICICI app, right? They're complex enough to basically navigate through, right? So that agent can sit on top of that app and guide the customer complete journeys. Do do stuff in that app. So that's one is second. Uh, second is a second is a new form factor that we have created. It's, uh, it's an avatar sort of an app, uh, agent, sorry. Um, which has generative UI on top of it. So you just talk, the surface remains the same. It's just that elements are changing.
Utsav Somani: Interesting. And what kind of permissions do you require? Is it mostly app based or app level or does it involve OS level permissions as well?
Ashutosh Prakash Singh (RevRag.AI): No, it is majorly app level permissions that we require.
Utsav Somani: And banks and BFSI apps are willing to give you that permission that easily?
Ashutosh Prakash Singh (RevRag.AI): So we are, uh, basically an SDK on inside their app, right? For a user level, we only require a microphone permission, not anything else. Right. And, uh, when we are embedded in their app, uh, the data is with the bank or with the FinTech and everything, right? So it's very, very compliant from that lens. And at the same time, uh, when we are doing and basically the agent is chatting or agent is guiding and doing stuff, it is all PII redacted, uh, from, from the, uh, from the conversational point of view as well.
Dhruv Sharma: And so some of these apps that we're using the examples of Ashutosh, let's say the ICICI net banking app, there's a hundred things it can do, right? And now are you talking here about a general purpose agent can, that can do all of those a hundred things equally well, or are you putting at the general purpose agents disposal, like a fleet of these specialized sub agents that can do 10 things well, and 10 things well, and five things well, and so on and so forth.
Ashutosh Prakash Singh (RevRag.AI): Right. So, so for, uh, so for a very big on the pricey app, right. Uh, it's not easy for anyone to get inside that app, first of all, and, uh, doing that general stuff, because if let's say an ICICI is there, right. The P the personal loan team works differently. The let's say insurance team works differently, right? So you have to combine all of them together and do basically, if you want to do general, then you have to bring everyone on the same page, how we do GTM is that we go to, let's say, uh, uh, personal loan app, right. And, um, that journey we first saw once there is a hit overall, then we get to the homepage and start the whole navigation. What we have started doing is that first we show, uh, one, one, one journey, which is like, let's say is able to complete the agent is able to complete the whole journey that we ask for. Let's say, because this is working out for your users. Why not to put it on the homepage as well, right? Or even let's say a redesign, the overall experience with a conversational point of view. So we have done this for two to three apps today and slowly and gradually we go inside these BFSI because we are very, very BFSI focused. And the thesis is, uh, that see for a person like me, I can have a person come to my home and complete that home loan journey or complete that PL, uh, even create a savings account, but for a person who, let's say, uh, a merchant or a farmer, right, they don't have access to all of these relationship managers. Right. So for them to even get that loan, it's not easy for them, right? So India has like 27% of financial literacy, but after COVID we were pushed to adopt digital, right? So there is a problem even after a point, whenever a user is actually using all of these apps and there, these are complex apps, as you were rightly saying that, uh, even in ICICI, we have a lot, many products, right? So if I want to have an insurance, even in insurance, there are multiple insurances, right? And with different, different claim ratios, different, different, uh, uh, nitty-gritty is overall.
Dhruv Sharma: We've covered, in fact, covered the theme on the show before, which is, um, you know, from a customer service standpoint, AI can be a great equalizer. You can be sitting in a tier two, tier three town, maybe even a small village and get almost like a private banking, like experience just because AI can deliver it to you, you know, you don't have to be a, be a big account holder or anything.
Ashutosh Prakash Singh (RevRag.AI): That's what, that's what we are actually trying to do. And that's why, if you see, when I started, I told you about two agents, one is basically a co-pilot, which is on your app. The second form factor, which we are experimenting and it is, it is in experiment mode. And a lot of companies are actually adopting is that there is a face in front of you, which is an agent, and then things are changing on the surface, which is basically conversational in a way, right? Whatever you talk, that agent is able to fill it up. So it is a private banking that, you know, there's a lot of, I have this observation.
Dhruv Sharma: You remember like demos, some 10 years ago or pitches, right? People used to say in three clicks and then five years ago, they used to say in a single click and now it's zero click because everything's become conversational. And so even when you see pitches or like demos, it's just changed.
Utsav Somani: A lot of people are actually betting on the fact that the next set of interactions that you do with your computer in the future will actually be voiceless. That's what OpenAI is creating with that device, right? With that Apple designer.
Dhruv Sharma: Have you configured a dot yet, by the way? I have not, I've actually not received access to it. Okay.
Utsav Somani: All right. Have you?
Dhruv Sharma: I have, I call it something I'll show it to you sometime.
Utsav Somani: But apparently instinct is losing out because the connectors are going a little bit off and they're getting blocked in a bunch of places, but they're raising it 40 billion, which is just insane.
Dhruv Sharma: Yeah.
Utsav Somani: So I mean, this, I mean, yeah, Claude do it and a bunch of others. And I mean, yeah.
Dhruv Sharma: Personalize agents.
Utsav Somani: Yeah. But Ashutosh talk to us about the models that you're using voice. We've had so many voice AI founders on the show and many of them targeting the BFSI industry as well. Mostly the call center recovery and other mechanisms, other applications of that. But how are you thinking about your own models or working with other partners who are providing you these models?
Ashutosh Prakash Singh (RevRag.AI): Right. So we, when we started, like we didn't have an option to, we started using outside models, right. With the API and all right. But soon after we realized that for something, what we are building, there is no way we are able to get the right accuracy. So what we do is it is conversational in a way. And at the same time, it is action in the app. So it's, it's, it's basically a multimodal element to it. Right. So, uh, we tried using OpenAI. We tried using even cloud. We experimented initially, but it was, it was working out for a few instances, but soon it became very, very expensive for us to basically continue that way. Right. So we, we got out and in our, in our head, there was a mental block that we can't build a model. It's building a model. It takes a lot of effort, a lot of energy. We were like 10 people team. We didn't know what to do. So, and so forth. So what, but we took our, uh, attempt and basically started fine tuning a model, right? Uh, right now, what we have done is we have fine tuned a model, which is conversational AI plus action intelligence led, which can actually talk and take action, understand the whole context on the screen, guide the user, understand the context from the outside world as well. And basically, uh, uh, do, do the stuff inside the app and slowly and gradually, uh, what we would like to do, uh, is build our own model, but as of now we are fine tuning, uh, yes.
Dhruv Sharma: That's a good, I guess, uh, my final question for you, uh, take us with you. So inside an early stage funded company, what's the energy in the team? Like right now, what are the next couple of months looking like, uh, what are you guys excited about? What is scaring you guys?
Ashutosh Prakash Singh (RevRag.AI): Yeah, for, for us, uh, we were a very small team and we didn't raise for a, for a long time, we took a pivot, uh, last year. So we were building a fintech, we built agent for ourselves and then started selling the same agents to others. So, right. So we pivoted into RevRag.AI and, uh, uh, since then we have been fortunate to have worked with the top fintechs in India, top, top NBFCs, banks, one and so forth. So, um, and in last year, it was not easy for us to settle these in our pages because nobody was able to understand how they will work, right? But this year it is better and people are able to understand people are willing to take that bet. What we are excited about is building all of those models, scaling them now, now, because we are out of pilot phase to actually going live with scale, right? So that is something that we are really excited about. And we want to give this agent tech experience to all of the Indians that are out there and they should be able to talk to an alien.
Dhruv Sharma: Yeah. I hate to interrupt, but I think what's standing out and what you're saying Ashutosh is that maybe, you know, the thing to do over here is to be maybe six or eight months ahead of everyone else before enterprises are even ready to adopt the thing. And by the time you've built it, ready to put it into their hands, they warmed up to the idea of, you know, putting it in front of their users and, and rinse and repeat.
Ashutosh Prakash Singh (RevRag.AI): So what we have been able to understand by, but like, so what happened last year, it, nobody understood what is an in-app agent will do, like what is an in-app agent, right? Or what will you do? Uh, but after they have deployed voice agents, right. They've understood. Okay. Okay. This is something that you'll be able to do. Now they're able to comprehend better. What is an in-app agent? What is an avatar agent? What, what actually you're referring to, right now they're ready for us. And, and we were ready last year itself, but yeah, now going at scale. So it's, you're right. Uh, uh, majorly, and one more thing that you guys were, while you were mentioning about, uh, personal assistant, uh, assistants as well, right. So what we've realized and what we're seeing, I don't know, like, whether it will be true after two to three years or not, but what we are seeing is that all of these banks app will not give access to a dot or a instinct of sort, right? So there will be an agent which is sitting inside the app, which will be talking to all of these agents to let's say, get stuff done, right. For, for that user. So I think that is where we also see our, our play improving overall in long form.
Utsav Somani: All right. Ashutosh, thank you so much for coming on our show. All the best.
Ashutosh Prakash Singh (RevRag.AI): Thank you, sir. Thank you. Cheers.
Utsav Somani: All right, listeners. That's a good note to end the show on. Hope you've had fun with us this Friday. We'll see you on Monday. Same time, same place. Bye bye. Thank you. See you on Monday.