#62 Synthetic Human as a Service | Krishna (Vasanth) Namasivayam

Speaker A: What's the most frugal thing you've done? What? The hack that saved you money.

Speaker B: The hack that saved me money.

Speaker A: I'm your host, Vishesh, building Neo voice control for your entire computer. No typing, no clicking. Curious. Check out Vishesh Space. My next guest, Krishna Namasivam, um, is working on what he calls synthetic human as a service. After leading AI at Dropbox and Meta, he's now building with the radical belief that we should simulate human behavior before we ship code, product messaging, or anything really. In this episode, we break down his contrarian thinking on product AI hiring and what it takes to launch in a greenfield market. Hey. Hi, Krishna. Thanks for joining me on the podcast.

Speaker B: Thank you so much for inviting me.

Speaker A: Awesome. We've been talking for a year now. Uh, I think a few people ask me, you know, why are you doing this? And, uh, I think this is a perfect moment to answer that, probably, which is I love talking to founders and I love talking about challenges running, running a company. And it's been great talking to you so far and thanks for, uh, finally, uh, us getting to it. Since we spoke last year, things have shifted a little bit in terms of, um, where the product is and what you guys are doing. Do you want to talk a little bit about Featurely and what you guys are doing and ah, what are you addressing for your customers?

Speaker B: Yeah, absolutely. So at the very core, Featurely helps you simulate human behavior and you could imagine that you solve human behavior at scale to help you make better decisions as you're building end to end experiences for people. If I need to kind of look back, why did we start Featurely? We believe that we are living in a world or we are starting to live in a world, ah, at the marginal cost of building something is becoming lower and lower. Um, and so the bottleneck was shifting from can I build something? To is this the right thing to build? That became more and more obvious as we started, um, looking at how the market's playing out and as we looked further into the future. I believe in this world state where experiences are really tailored to every single person as they kind of interact the digital experiences around them. That's where I see the world going towards over the next decade or so. And so then the challenge was, well, I could theoretically generate a, uh, different experience, an end to end experience. And when I say experience in the broadest concept of the term, from messaging to all the way to the actual workflows itself with all the generative goodness which has come out in the recent years, but the bottleneck still remains. Do I really understand what the person on the other side wants? How do you kind of give them the it just works kind of experience, borrowing the slogan from Dropbox, where a person, the experience adapts to the person as opposed to vice versa.

Speaker A: Hmm.

Speaker B: And so then, you know, that's kind of the core problem we start off to solve in Featurely, and that problem was being solved by the concept of can I simulate how a human who I'm building for would value this experience, react to this experience, engage with this experience, and so on. Right. And that comes under human behavioral simulation. So that's kind of what Featurely has been about. Now, as you can imagine, the concept of human behavioral simulation, while relatively new, could potentially be put into a, ah, multitude of applications. And in reality, which application makes sense depends upon the demand for that particular use case. But B, the quality of the human behavioral simulation and what are the takes of the decisions being made based upon, you know, the simulated responses you're getting. Right. Um, so we have been kind of exploring the entire, uh, like that particular space and kind of triangulating into, you know, where we are most valuable. Yes, that's kind of what Featurely has been about.

Speaker A: Interesting. And are there any specific experiences that you've already kind of solved for and started seeing results?

Speaker B: Yeah, absolutely. So I think there are two broad workflows or use cases which we are solving for today. One is on the digital product side of the world where, uh, we work with especially, but not limited to vertical SaaS, companies which are building in markets where their end users are not easy to access. They're kind of unreachable example constructs, in some cases compliance.

Speaker A: Right.

Speaker B: Donor platforms on higher education, many of these kind of products, your end user is almost unreachable. And when your end user is almost unreachable, it's hard for you to kind of fathom what is the right thing to build and how do you go about building it. Right. And so when you don't have an alternative, Featurely is a great fit to kind of come in and you kind of spin up Personas, you spin up synthetic humans within Personas, and then you chat with them, you get them to try your prototype, you get them to experience your prototype, you watch how they make decisions, and you use that knowledge to build a more seamless version of your product before you hit market. So that's one entire set of use cases which we're currently solving for. Second, uh, set of use cases which has been coming up more and more recently is on the messaging and let's say re onboarding side of the world. So it's more on the marketing side of the world where uh, you're kind of trying to figure out how do you increase awareness, what is the right messaging which triggers the right response in your target audience. And so this is more in the awareness and acquisition part of the funnel. And this could be landing pages, it could be emails, it could be other pieces of content, ad creatives, things of that sort. So we do see that as well. But again in those vertical markets where the end users are very difficult to reach out to and you said you

Speaker A: allow people to even watch these synthetic users sort of use the product. How do you do that?

Speaker B: Yeah, absolutely. So like taking a step back, one of the things which building products before featurely ethnographic studies where you get to watch a user interact with your platform, with your product and always garners gives you a lot of information. So we try to create that same experience in the digital world. And that implies, let's say I'm building a product for, I don't know, um, construction four men. So that's a Persona, right? And uh, while that's a Persona, it's not a heterogeneous archetype. You could have, could have diversity within the Persona. So we create synthetic humans of various demographic, geographic, psychographic attributes and behavioral attributes. They come in they your prototype or your platform and you can give them a workflow and they will go about, you know, in an autonomous manner figuring out the workflow. As they figure it out, you have a video captured of what they're doing and at the same time they're talking out loud in terms of what's working and what's not working. And you can always ask follow up questions either through the inbuilt user, researcher, agent or manually as well to kind of, you know, poke and prod and get as much information as possible. So I've found those experiences to be extremely valuable from a knowledge perspective. And you know, that's kind of what we are trying to mimic in the synthetic human world.

Speaker A: Yeah, that's super cool. And are you dog fooding this for yourself?

Speaker B: Absolutely. I think we, our entire team, so we don't have a quote, um, unquote product manager. Our entire team is feature like our engineering team uses featurely to kind of figure out what to do on the next sprint where they keep getting synthetic humans to jump onto the product or the landing page or the messaging or whatnot and you know, test the market, test the users Give us feedback and then we use it to go to the next thing, so on and so forth. And so that's kind of balanced the cycles between user research and deployment.

Speaker A: And what are the chances of the simulation being of going off the rails? Right. Giving you feedback that is completely, uh, relevant or wrong even?

Speaker B: Yeah. So there's definitely a non zero chance of that happening. What we have seen empirically with customers so far is the likeness score between our synthetic humans and real humans of that particular type ranges in the E to 85% and in some cases up to 90% in terms of human likeness. Now that said, you know, is there a possibility. Yes, there's always a possibility.

Speaker A: Right.

Speaker B: But I think this is where there's human judgment involved, where you kind of ensure that you discard data points which are completely off. And for what it's worth, Vishesh, uh, I'd like to point out that this is not new, like even in the pre synthetic research world.

Speaker A: Right.

Speaker B: When you create user research panels, when you do surveys every now and then you get responses which are like way, way, way out somewhere, which the human has to decide whether, you know, it is a legitimate response or if it's not. Right. So it's very similar to that. You'll always get these outliers and which is why you need a human in the loop.

Speaker A: Yeah, that's true. I mean, when talking to users, sometimes you really have to decide feedback is worth taking into account or it has some sort of bias. Right. And before this, you know, you've spent time at Dropbox Meta Nvidia. What made you take that leap? And, um, how are things so far as a founder?

Speaker B: Yeah, so that was the first question. What made you take the leap? I think at the very core I was running into the problem Featurely is trying to solve on a daily basis. And you know, um, it was personal enough where I was like, I want to solve this problem for myself. And then the, the, the, the seed of the idea started coming out of that point. Secondly, you know, it felt like the perfect time in terms of how the macroeconomics was aligning. What was possible with compute today was not possible 10 years ago. And so it made a lot of sense to it. It was like, well, I have a problem. You had a reasonable idea how to solve it. And unlike 10 years ago, we had the compute available to, you know, do complex things like human behavioral simulation. And so, you know, those three things came together and it kind of gave me the impetus, uh, to move on from big tech and you know, uh, get to startups. Now that said to your second question, how's life been as a startup founder? I think one phrase, it's been a roller coaster, it's adrenaline filled, it's a lot of fun. Also, you know, there's a lot of ups and downs. You learn like crazy. You learn things which you never thought you would ever learn in your life. Uh, so that is a lot of fun. But yeah, uh, unlike the, when you're in big tech, you are kind of working with an almost infinite cash engine in terms of what you can do and what you can't. And so one of the very first things you learn in this journey is being frugal and kind of doing things the prudent way as opposed to just blowing stuff on things. So yeah, it's been a lot of fun.

Speaker A: What's the most frugal thing you've done? What? The hack that saved you money.

Speaker B: A hack that saved me money. I think it's a bunch of things. But I guess to begin with it was something as simple as the way we went about hiring.

Speaker A: Mhm.

Speaker B: Didn't retain any recruiting agency or anything of that sort, um, which a lot of founders based in the US when they're dealing with India recruiting tend to do. We instead were spending a lot of time initially hiring through Reddit or through LinkedIn, you know, um, and yeah, with very interesting results either way. But as I think about it, that was probably the very first thing which we did, which we took a different than what you would consider more.

Speaker A: And so how did, like, how does hiring through Reddit work? Like, what were you looking for? Did you specifically go in Reddit looking for talent or did that happen by chance?

Speaker B: No. So my thesis was LinkedIn has become a very performative platform where what you see is not what you get. Now Reddit, because of the anonymity, people tend to be more themselves.

Speaker A: Mhm.

Speaker B: There are some pretty strong communities in Reddit related to products, startups and whatnot. And so I started hanging about a lot in those places, looking for potential matches who'd be willing to build featurely alongside me. And that's kind of how it was. That's how it all started. The thesis. It definitely was more, uh, what's the right word? Authentic Than trying to hire through LinkedIn. It was less performative and um, we did make a few hires. I think the first person I hired through Reddit is still Apollo Featurely. So is this a, ah, perfect solution in any way? Absolutely not. Just like any other hiring mechanism, you have the successes and you have not so good outcomes. But again, as a founder, you're constantly trying to see what you can do to improve your shots at success within the constraints you're dealing with.

Speaker A: Yep. Are you still hiring from Reddit? And um, you know, now that it's been a few months, what, what's some, what are some of the things that people should and shouldn't do while hiring from channels like these?

Speaker B: Yeah. So am I still hiring from Reddit? Less frequently than earlier. What ends up happening is as you start building a critical mass in the organization, people start, uh, attracting their own peers from their past lives and therefore the hiring challenge kind of evolves over time. I think Reddit was a great, oh, it was a very interesting hiring strategy when you're starting at the seat and you're starting from zero. But obviously, like, as you get more and more people, your channels to access talent kind of changes a little bit and so the balance kind of changes. In terms of the do's and don'ts of hiring through Reddit, I really don't think there's a playbook as such. It's about being really embedded in these communities, which, for example, our startups or RSD or you know, developers, whatever region you're focused on and kind of understanding what the market looks like, what people are looking for. Uh, fundamentally we change, similar to building a product. It's about trying to understand the market as a whole. So you're really trying to understand who are the people you are attracting and you are observing them and interacting with them in a space where people are truer version of themselves as opposed to LinkedIn. And in an ideal world, that's how you build great products and that's how you kind of try to attract the right talent.

Speaker A: Yeah, makes sense. Yeah. I think the point on hiring, once you have critical mass, the hiring channels change. Hiring experienced talent actually adds a lot of value in hiring future talent because, you know, they have a network of people, um, that they can bring in. So always saving money is not good.

Speaker B: Absolutely, absolutely. And those are things which run along the route. Right. I mean, I think you need to have a good mix of people. It's about experience, it's about talent, it's about skills. And especially for startups, it's also about do folk have clarity on what they're signing up for. The challenges with, especially tech startups and AI startups, is there's a lot of hype around them. And at time the concept of working in an AI startup is more attractive than actually working in A startup at large. Right. And at times, because people are blinded by the bright lights, they sometimes lack clarity in terms of what they're leaving on the table when they choose to go down the startup path. Inevitably, when the clarity doesn't exist, it leads to mismatched expectations and, you know, heartburn on either side or both sides in most cases. But yeah, those are very important things to kind of look out for. Yeah, that's uh, hiding from whatever channels, including Reddit.

Speaker A: And so you've said that problem is in design or code. And I think almost every founder agrees it's building someone, uh, building something no one wants. You know, why do you think the current product process, where is it breaking? And why are more and more people building stuff, uh, that no one wants?

Speaker B: Let me kind of look at this from two different perspectives. Let me look at it from the perspective of larger established technology companies versus small, like two member teams trying to figure things out. In larger established tech companies, the incentive structure for engineering product design teams is not always aligned with shipping and experience which delights the customer. So firstly, there's an incentive issue. Secondly, apart from the incentive issue, there's a practical issue of being able to deliver goals every quarter and not having the time or bandwidth to kind of do a very robust evaluation of the market before, you know, building what you need to build. So that's a second like, like it's a very structural issue which, uh, has been there for a long time. And then along with that there's an organizational issue where more than we'd like to admit, a lot of building happens based upon gut instinct. Loudest voice in the room, highest position. Like what? Like you take your pick because you want to move fast. You don't have time to, um, do the necessary due to the rules. And so you kind of start doing this. And so that's kind of what happens today or until recently in most large technology companies or established companies when it comes to. And so with the combination of all this, not really understanding your end customers was almost status quo. Everybody knew that. You know, we had holes in our entire discovery process, but the concept was, well, that's okay, let's compensate for it by doing a lot of AB testing, by moving fast, releasing a lot of things to the market and being generally agile. And that was a fundamental thesis. And I think it's worked to some extent what it's worth. But now we live in a more, uh, generative world. And so some of the dynamics or the build side of the entire process is kind of shifting and so then, um, you know, um, we might need to reconsider how we accepted the status quo, uh, in the earlier versions of product building. Now when it comes to really small companies, the other segment which we were talking about, I think small companies or like, I don't know if you're a two, two member team or a five member team with very, very limited financial resources, you are going headlong to somehow breakthrough. And if you tell me that, oh, I need to do customer research for eight weeks, uh, before building something, well, the physics of that doesn't work out in a two or five member team. This is kind of looked at as an afterthought as opposed to what should be done. And I mean like statistically this kind of backed right across the world across the last decades, somewhere between 60 to 80% of any experience built is either never used or rarely used by the end customer. And uh, you see the statistics from Pendo and from so many other places. So statistically this is what has been happening in the world and we've kind of been accepting that statistic because there wasn't a better solution, very, very frankly, which would fit within the budgetary time, logistical and organizational constraints which existed. Now I think things are shifting a little bit because just like how the generative technology has kind of changed how we build stuff, people are recognizing that we can build stuff really fast. But if you keep putting out stuff, I mean, like, which means that at any point of time you could have a thousand ways or a thousand ideas or what to build. And so that. And so then even if you can build thousand things really fast, at the end of the day you're selling it to a human on the other side. And the human's attention span is finite. So you do need to filter down from your thousand idea list to, I don't know, a far smaller number of and possible ways to do things. And for that you'd want to pretest against simulating against humans.

Speaker A: Yeah, I kind of see that problem already with some of the products that are shipping like at breakneck speed. Like if I'm heavily using Claude and Claude product like anthropic products. And yeah, they are releasing every day the product is completely, if not completely shifting, shifting by 20%, which is still a lot for a user to take in every day. You said you, at this point you don't have a pm. How has the team structure shifted and where are you seeing success yourself?

Speaker B: One thing which I kind of like to point out is the PM as a title is a relatively recent development when you look at the overall life cycle of building software, um, I might be wrong here and please do fact check me, but I think sun existed for the longest time microsystems without the concept of a, ah, product Manager. Now that's not to say that the work being done by these titles was not being done. It was just organized, uh, organizationally done very differently. But you still had someone making decisions on what is getting built. You still had someone coordinating across teams, you still had someone really trying to understand the market and whatnot. But that person or that group of people were not called product managers, at least in the earlier versions of the tech industry. Um, but I guess with Microsoft I think they used to call them Program Manager and then Google, uh, I think was very instrumental in making the product manager title extremely sought after in the early 2000s. Um, with that historical context I again see something similar happening. Regardless whether there is a title called PM or not, at the end of the day somebody is responsible or some group of people is responsible for making decisions on what's getting shipped, what is wrong, what needs to get built, things of that sort. Some group of people are going to be responsible for debugging stuff, evaluating what is going wrong in the funnels and fixing it. Some group of people are going to be responsible for really understanding the market. What I see is because of the advent of many of these tools, more of these functions can be done to some extent by a single person as opposed to specialized titles for each of these uh, things. And that's just because some of these AI tools can help you do so much more even with a smaller team. And so this is directly applicable to startups and founder teams where I see that in a bid to extend Runway and in a bit to be lean, you want to, a lot of folk hire the minimum possible set and they augment themselves with a plethora of AI tools across workflows, including on the product side of the house, the how do

Speaker A: you ensure the quality of the decision is at the level of, you know, for example, you know, if you let engineers take that call, but if they have no experience dealing with those kind of decisions, have you, have you seen impact of that happening?

Speaker B: Yeah. So for what it's worth, in um, the startups sector, which we're looking at historically as well, most of the product decisions were generally made by the founding team or some group of members in the founding team. And I think that still continues. It's just that they use some of these tools to augment them or to make them make better decisions. As they go about doing so now, uh, even on the engineering side of the world, I mean I believe that most people can pick up skills outside their traditional strengths as and when required. And so I think uh, tools like these help them develop those skills. So if you go back in time, right, when product like V0 first came out, yeah, it could be used for design. But the people who were using V0 most was not necessarily designers, it was engineers and product managers who wanted to spin up quick prototypes to kind of showcase what they're trying to build. When Cursor first came out, I remember so many peers who were non technical in nature were very excited about the fact that they could build a startup using Cursor, even though they're not very technical. So you could see that sometimes these AI tools are used by people who don't necessarily have that skill set, but who need that skill set to kind of fill their, that hole, you know, until, or you know, alongside getting a human to kind of play a part in the entire thing.

Speaker A: Yeah, finding someone with the expertise and then getting them to work on what you want is, is obviously.

Speaker B: Yeah. I mean if you go to Reddit startup, there's a graveyard of people who claim they have a great idea and they want a technical co founder to build it for them.

Speaker A: Right.

Speaker B: Like that's almost become a um, it's a meme at this point of time. And you know, just because a person is technical doesn't mean they cannot come up with their own ideas. They probably have their own passions they want to build for and you know, like whatnot. So I think, you know, in a weird way these tools have kind of enabled more people to dream of building, you know, a product which they could not have in a uh, non generative world where they were limited by their skill sets one way or the other.

Speaker A: And so for you, what's an ideal startup team structure? Like what's the composition looking?

Speaker B: It really depends upon the company or what you're trying to build, at least feature. We are pretty heavy on the machine learning side of the house, uh, as we are building some of our own models and quite some on the full stack engineering side of the house. We are very light on design and product and those kind of functions. So the founding team kind of manages it among them.

Speaker A: The uh, product is pretty unique. What has been the GTM journey for you so far? Are there any lessons there?

Speaker B: What's been different? So this is where it comes down to category defining. Like whenever you come up with something which is potentially category defining, there's quite some engines in the world. And like honestly for Featurely and for companies in the. When you go up to a person and say oh, I can simulate human behavior, reactions range from wow, that sounds like science fiction to I don't believe it. And so for products in this stage you both need to allow people to try out the product in a lightweight manner but have a, some kind of sales assist function to help them really learn the product and gain maximum value from the product. So I would say our uh, go to market has, it's basically product led with sales assist, you know, for the plurality of our design partners. Now as we deal with larger companies, of course you do need um, effective sales um, uh, system to kind of navigate selling to larger companies. And I think this is the same GTM or this is similar GTM process used by many of the companies which have come out in the last few years. Like I think about cursor and I think about V0 and I think about so many of these tools. You could try a version of the product at almost zero cost but a very, very limited version of the product. And then you kind of have people kind of help, uh, I mean uh, in some cases helping you kind of onboard onto the product. And then you know, as you get, become a more serious user, you know, and as a company becomes larger they explore more dedicated sales functions to kind of work with larger companies as assisted.

Speaker A: Are there things that you've done to uh, let users experience that time to value faster? Are there any learnings there?

Speaker B: Yeah, so absolutely in these kind of things TTV is very important. So particularly in Julie's case, I think there are two things we experiment, experimented with. One was anyone could come into a um, let's call it a non logged in version of the product and then test their landing page against one synthetic human, get feedback which they can try to implement. But if they need more richer feedback, more humans, et cetera, they would need to log into the product and start purchasing credits. So that was definitely one of those things which we did second with Featurely. We kind of work both on the individual simulation side and the population side. So um, we also have a few digital twins available whom you can converse with um, without logging in for a very limited number ah, of sessions and limited number of people. But uh, if you find value you can obviously onboard in quicker uh, uh, into the entire process. Right. So what we figured was even before a person has to log in they would want to see some value in a matter of minutes and then that would Kind of, you know, uh, grease the funnel to kind of getting them to sign up for the product. Then as you sign up, we have a very specific kind of onboarding experience to get you to simultaneously explore the product, but also see your first result in a matter of minutes. And then at that point of time we open it. Because typically when you do complex simulations, you're talking about a reasonable amount of time and that kind of hurts with your trying to get people value really, really fast. Right. So you give a lighter, uh, weight version of the simulation which is not as accurate, but people can kind of feel the value. And then as they get more and more into the process, you start bumping up the simulation quality. But it also increases the time to simulate M. Interesting.

Speaker A: And on the messaging front, how did featurely help you test the messaging? Like did you from the get go dogfood featurely for messaging? Uh, and what are some of the experiments you've done on the messaging front?

Speaker B: Yeah, so to begin with we dog footed primary on the product side because that's how the feature product itself was built out. It wasn't built initially for messaging and marketing workflows, but then once we start building out those workflows more recently, we started dogfooding it against that as well. And so even like we very recently started working with you know, cold outbound and LinkedIn. And so there's a variety of tools we are using for that. But message, uh, what do you call adaption is kind of done through Featurely to go and hit those people. But we do it like in a B tested way. Right? We do through feature lead, but we also do it through other means and kind of see which way works better and things of that sort. And then we go.

Speaker A: So we've reached a section I uh, call never get a cold email and these are things that we'll highlight. Hopefully we'll let people get to know you better. A book that's been most helpful as a founder.

Speaker B: As a founder. So I think personally I think, let me kind of look at it from two perspectives. One from the perspective of just resilience. And I think I personally like this book by Viktor Frankl. It's called uh, Search for Meaning of Life. I think it's uh, a, it's a very special book and I know it's not related to entrepreneurship per se, but I think it's very, very um, it's a book which had a large impact on me. But suppose I look at exactly what we're doing today. I think um, you know, uh, there's a bunch of books, something like, um, it's a book by Nate Silver on, um, um, signal and noise. And I think that's very, very relevant to what we do at Futurely. And so that was very exciting.

Speaker A: I've not come across the signal versus Noise book, so I'll check it out for sure. Um, one advice you'd give every founder you met today, I think it's about

Speaker B: building every founder or every wannabe founder.

Speaker A: Every wannabe founder, because we don't want to judge. They could eventually be a founder.

Speaker B: No, I think looking back, let me rephrase this as what would I have told myself two years ago or a year and a half ago if I could go back in time? Um, I'd have probably told myself to start the journey with a lot of clarity of what your ultimate goal is. There's always a ton of distractions, especially in the startup space and the tech startup space and the AI tech startup space. Right. As you go further. And it is so easy to get caught up in all the noise and the distraction and the comparisons and whatnot. But it's very important to hold steady, uh, to what your eventual direction is and know that your path might be different from, you know, take any startup which is currently in the headline today and not panic about that, you know, so, so, so that is something which, um, is kind of what I would. If I could go back in time. That's kind of what I would tell myself.

Speaker A: Clarity on the mission is so important. Um, I think it was a recent tweet by Elon Musk, and I don't agree with a lot of, um, everything he's doing, but one of the tweet was, uh, the goal is to understand how the universe works and everything flows from it, or something like that. And then he broke down that. And it was a very simple tweet, but very profound. As a startup, if you have your mission nailed down extremely clearly and with a lot of clarity and crispness, everybody in the team understands and then they're all aligned. Cool. Really enjoyed our conversation. Where can people find more about, um, Featurely? And are you guys hiring?

Speaker B: Yeah, uh, you know, featurely AI. Featurely AI is where everything is. And follow me on LinkedIn because you get the most updates on what's happening Featurely. What are the, uh, new boundaries you're pushing in human behavioral simulation and what possible applications you can take advantage of. Generally comes out as posts in LinkedIn. So find me on LinkedIn and hang with me. I'd love to chat with you. Thank you so much for this. It was a really delightful evening talking to you. Um, thank you.

Speaker A: If you found this valuable, you can subscribe to the show on Apple Podcasts, Spotify, or whatever your favorite podcast app is. See you in the next episode.