In this episode of 'The Next Next', host Jason Jacobs interviews Swupnil Sahai, the co-founder and CEO of SwingVision, an AI-driven app for tennis and racket sports performance tracking. SwingVision utilizes a single iPhone or iPad to provide real-time stats, automated line calling, and video highlights. Swupnil shares his journey from playing tennis and working at Tesla's autopilot team, to creating SwingVision by combining his passion for tennis with AI technology. The discussion covers the early challenges, data collection methods, and model training for SwingVision's development. They also explore the app's go-to-market strategy, customer acquisition, and future plans, including expanding to other racket sports and exploring new technologies for reduced user friction. Additionally, Swupnil offers insights on capital efficiency, maintaining flexibility, and the importance of entrepreneurship in leveraging emerging AI tools.
Building SwingVision: An AI-Powered Revolution in Racket Sports
In this episode of The Next Next, host Jason Jacobs interviews Swupnil Sahai, co-founder and CEO of SwingVision, an AI-driven app for tennis and racket sport performance tracking. Swupnil shares his journey from working on Tesla's autopilot team to founding SwingVision, blending his passion for tennis with his expertise in AI. The app leverages a single iPhone or iPad to provide amateur players with pro-level analytics, including real-time stats, automatic line calling, and video highlights. They discuss the inception of SwingVision, early challenges, data collection strategies, and the importance of user feedback in refining the technology. Swupnil also talks about the company's go-to-market strategy, funding journey, and future aspirations, including potential expansion into other sports and improving infrastructure at sports facilities. The conversation highlights the merging paths of entrepreneurship and AI in sports tech, providing valuable insights for aspiring founders.
00:00 Introduction to The Next Next
00:33 Meet Swupnil Sahai: From Tesla to SwingVision
01:37 SwingVision: Revolutionizing Tennis Analytics
04:16 The Reality of Entrepreneurship
06:49 Swupnil's Journey: From Academia to Tesla
11:15 Building SwingVision: The Early Days
14:47 Training AI Models for Tennis
19:40 Gathering Data and Scaling Up
22:07 Launching the MVP and Iterating
23:27 Advice for Aspiring AI Entrepreneurs
25:15 Customer Feedback and Data Utilization
26:56 Fundraising Journey and Initial Challenges
29:01 Go-to-Market Strategy and Organic Growth
31:23 Adapting to AI Advancements
36:03 Freemium Model and Product Experiments
41:23 Future Vision and Expansion Plans
49:01 Final Thoughts and Encouragement for Entrepreneurs
Jason Jacobs: [00:00:00] On today's episode of The Next Next, our guest is Swupnil Sahai, co-founder and CEO of SwingVision, an AI powered tennis and racket sport performance tracking app. Designed to make professional level analytics accessible to amateur players. The app utilizes a single iPhone or iPad to provide real time stats.
Automated line calling and video highlights eliminating the need for expensive multi-camera systems. Now, Swupnil's story is interesting. He played tennis. Growing up and he observed how, , if you look at the pros, for example, they have access to all kinds of fancy stats, but none of that trickles down to amateur or emerging players.
And as an amateur or emerging player himself swept no one of those kinds of stats for his own game. And meanwhile he was working at Tesla on the autopilot team. He got [00:01:00] there, I think he was one of the first six people on that team, and he was developing AI for 3D object tracking. On that team and the combination of his wants as a tennis player and his experience at Tesla led him to combine the two and build SwingVision.
Now, we cover a lot of ground in this episode and Swupnil's very transparent in terms of some of the early learnings, how he went about it, how they went about training the models where he believes. Lock in or defensibility comes as the landscape continues to get more crowded. We talk about SwingVision's strategy, who they serve, who they aspire to serve over time, and how they see the company evolving.
We talk about their go to market, what's worked, what hasn't worked, what he's learned along the way, and what muscles they're trying to build in that regard going forward. And we also talk about capital efficiency, venture capital. Staying narrow and [00:02:00] focused versus expanding. How to know when to be narrow, and how to go broad and when to stay lean and when to raise boatloads of money.
So at any rate, as a potential aspiring entrepreneur myself in a related category around this whole sports ai, coaching, computer vision, et cetera, I learned a lot from this one, and I hope you do as well. But before we get started.
I'm Jason Jacobs, and this is The Next Next. It's not really a show, it's more of a learning journey to explore how founders can build ambitious companies while being present for family and not compromising flexibility and control, and also how emerging AI tools can assist with that. Each week we bring on guests who are at the tip of the spear on redefining how ambitious companies get built, and selfishly the goal is for this to help me better understand how to do that myself.
While bringing all of you along for the ride, [00:03:00] not sure where this is gonna go, but it's gonna be fun.
Okay, Swupnil Sahai, welcome to the show.
Swupnil Sahai: Hey Jason. Thanks for having me.
Jason Jacobs: Thanks for coming. I'm excited to have you on as we spoke a little bit. I've been on this journey to figure out how to build different with a small team and more flexibility and control and looking at how AI can help. And now I'm turning my interest to. Maybe doing something in sports and fitness, getting back to my roots.
And it turns out you've been, sports and fitness is a big world, but within the racket sports world, you've been quietly building up quite a following. And I'm, yeah, I'm excited to have you on and learn more about what you're up to and and dig in. So I really appreciate you making the time to do
Swupnil Sahai: oh, of course. It's like pleasure's mine to be here. We've always looked up to. Runkeeper that you started. And so it's yeah, it's a, it's been a journey as as you're familiar [00:04:00] with. It's challenging, but I think yeah, we've been just pushing along, I guess quietly, but some people know about us more than others, but it's still, I feel like some people don't know about SwingVision still.
So it's an exciting time. It still feels so early for us.
Jason Jacobs: Yeah. No, but I can hear it in your voice. I think one of the. One of the big myths of entrepreneurship is is that it's all confetti and main stages at big events and fancy investors and private jets and, everyone kissing your ass or whatever. But actually, like the vast majority of it is just curled up in fetal position try trying to get the computer to stop running so you can get a few hours of sleep before you get up and.
And and start the process all over again the next day.
Swupnil Sahai: Yeah, a hundred percent. I feel like people don't really know it until they get into it. And your perspective of a founder is so different until you actually live it. And it's just it's a lot more pain and suffering than what it looks like.
Jason Jacobs: So is SwingVision your first the first company that you've started?
Swupnil Sahai: Yeah, actually it is the first company I started. Yeah, I've, it's weird 'cause I was always really risk
Jason Jacobs: I.
Swupnil Sahai: growing up. [00:05:00] Like I always liked to do things by the book and I went to college and then I went to grad school 'cause I was just like the safe path and then I was like an engineer after that.
So I was just, I never thought that my personality type, I'm also more introverted, I would say. So I think it was a bit weird for me to go down the founder path. But yeah, I just had this like burning desire to solve this problem and no one was doing it at a scalable way. And it's just like a personal, like problem I really needed to solve. I was like, oh, I think I can solve this. Let's take a shot at it. But yeah it's been weird. I don't think I'm like a traditional founder in some ways.
Jason Jacobs: I don't know if any founder's, a traditional founder,
Swupnil Sahai: true.
Jason Jacobs: if there is such a thing. Yeah. But I want, I definitely wanna get into the founding story and how the idea came about and all that. But just to frame the discussion, maybe just give a quick overview on what SwingVision is.
Swupnil Sahai: Yeah, so SwingVision, today's primarily a mobile app that you can download on your phone and you can use it to record your game. When you're playing tennis or pickleball, you set up your phone behind you on the court, and what's really cool is that we have our own AI that processes your video in real time.
And so it's tracking [00:06:00] the ball trajectory, the player movements, and so it'll generate. The same kinds of stats that you might see like on TV when you're watching Wimbledon or the US Open, like how fast is the ball, where is it landing? But then we also generate highlights of your best points. You can even challenge a line call.
If there's a close call that you're not sure about, you can use your Apple watch to check it. Yeah, we've added lots of functionality around that. But basically the idea is that it's bringing you the pro quality experience, but on any court, and it's using just the phone in your pocket.
Jason Jacobs: And, this is a technicality, but in order to do the line fault, so would you and your opponent both need the camera set up or the phone set up behind you?
Swupnil Sahai: No, what's incredible is just with one phone, we can actually see both sides of the net and
Jason Jacobs: I.
Swupnil Sahai: lines on both sides more accurately than the players at this point. So
Jason Jacobs: Wow.
Swupnil Sahai: Yeah. Yeah. And it's being used even in actual tournaments now in the us which is pretty exciting to see.
Jason Jacobs: Crazy. Now let's get into the origin story. So you were now, were you at Tesla most recently before starting? SwingVision.
Swupnil Sahai: Yeah, I was right before starting Scion, I was at Tesla. That was actually my first job out of [00:07:00] grad school. My first full-time job, I guess ever really.
Jason Jacobs: I.
Swupnil Sahai: I was on the autopilot team there at the very early stages. I was I joined the vision team when it was just like six people basically.
So I was an intern there. I. And in grad school, really enjoyed my time there. And after I finished my PhD, came back to California and worked on the autopilot team there. And obviously it's grown so much since then. But yeah, and I was specifically working on object tracking, so using the cameras to track the cars and the pedestrians using computer vision and neural networks to do that. and yeah, I joined there back in like 2017. Interned there in 2016.
Jason Jacobs: And so how many years were you there
Swupnil Sahai: I was there about a year and a half. I'd say. It felt longer, especially 'cause I interned also the summer before. But yeah, it was the total time, like full time was about a year and a half.
Jason Jacobs: and and what came first, this specific idea or the idea of becoming an entrepreneur? I.
Swupnil Sahai: I think definitely this specific idea came first. I actually thought of the concept for SwingVision [00:08:00] when I was in grad school. So I was, I went to Columbia for my PhD, so I was out in New York for four years and. Yeah, I've always played tennis my whole life. That's been my main sport. Also played a lot of basketball, so various ball sports I would say, but tennis was always my main one and did my PhD in statistics, so I'm just like a data person. And so anytime I'd watch professional sports, I'm always looking at the data and like really interested in the stats and all the reporting that they show on ES, espn now, it's ama, it's amazing, right? Like the kind of data they have, the professional level. And watching tennis, I was like, wow, I wish I knew some of this data about myself, how often is my first serve going in? How often is my second serve going in? What's the speed of my serve? I don't even know these like basic things. How often am I holding serve? Like all these stats are very ordinary if you're a professional player because you just take it for granted. It's all tracked.
But anybody who's not on like a broadcast match doesn't have access to any stats about this. This is like pretty basic information I feel like that every athlete should know about their game. Yeah, that was starting to come into my mind when I was in grad school. But I
Jason Jacobs: Wait. Can I ask one? Can I ask one, one clarifying question [00:09:00] there? Most of the pros, did they grow up like on the future pro track where they had access to all the stats growing up? So they don't know any other way or did most of the pros? Yeah.
Swupnil Sahai: bring up a good point.
Yeah. One of the first like conversations I had when we were raising money was with a former pro James Blake. He was like. He got as high as a number four in the world American player, and he actually brought this up too, which was when he was growing up and trying to go pro, he also didn't have access to his data.
Really only if you get to the point of being, let's say a top 100 player in the world and you're playing in like a. setting that's gonna be broadcast. That's the only type of scenario where they have the resources to pay for this kind of advanced technology. And at that level, it's super expensive.
It's like hundreds of thousands of dollars. It's like several cameras set up around the court triangulating the position of the ball to track all these stats. So it's a very complicated, expensive setup and like it only really make sense and. if it's a broadcast match and if it's a match that you're playing that's not broadcast, or if you're on the practice court [00:10:00] even, you typically don't have any data actually at all.
Jason Jacobs: And la. Last clarifying question, then we can get back to the founding story. But who are some of the players that are providing that expensive, hundreds of thousands of dollars service.
Swupnil Sahai: the one that's most familiar is Hawkeye. So they provide all of the ball tracking data for the Grand Slams for the A-T-P-W-T-A tournaments. They also operate beyond tennis. I think NFL is gonna start using Hawkeye this year actually for the 10 yard line and determining whether you actually made it to the line.
And so it's really cool. So they do really great tracking and officiating for lots of sports. And I think tennis is where they started, but they're, I think they're in like 20 sports now. But, it's a very expensive. Technology solution, it's obviously super accurate. But it's really geared towards like the highest level of athletes, the best in the world.
And so it's like super expensive, very accurate. And then what's interesting is like outside of that, there's like basically nothing. And they, some solutions have come and gone where they'd set up like. A few cameras on the court and it's like slightly as less expensive, but it still would be like several thousands of dollars of like hardware installation. And so no one had really built [00:11:00] anything that was like scalable in any form. That could be used by a high school athlete or an adult playing like out of their cl on the weekends. So that's really the market that we, that I felt like there's a big gap in and could be served and would like to be served with a solution like this.
Jason Jacobs: Okay, so you're working at Tesla, you're watching tennis on tv. You grew up playing tennis. You admired the stats on ESPN. You thought, why can't I have stats like this about myself? Then what?
Swupnil Sahai: Yeah looked at what was the existing solution. Obviously Hawkeye we talked about already. There were a few other that were, players in the space where they would install fixed cameras. But it's like a lot of cameras still really expensive, and so I. I think working at Tesla, working closely with Elon as well, like I was in a position there where I was actually presenting to him every week, and so we had lots of like discussions and I think one thing that was, I just learned really well from him was just like thinking from first principles. And a lot of the problems I was working on was around like cameras and like tracking how far away the cars are. And that's really hard to do with just a camera. Traditionally you need like at least two cameras. To [00:12:00] accurate, accurately triangulate like how far this object is. And so that's why you see companies like Waymo, they have additional sensors like Lidar or radar to help figure out. But it was interesting 'cause Elon was always pushing us to just do the vision, like just use cameras because fundamentally like humans can drive with just their eyes. You don't need like these fancy sensors to determine like exactly how far away this car is. It was interesting like intuition that he had and then. I always tried to, I actually at one point had a conversation asking him if we could add radar sensors to the back of the car because looking back there was no sensors other than cameras. And we had to solve this problem of tracking cars in adjacent lanes and like how fast are coming and can you merge or not. and so yeah, he basically pushed back, was like, okay, are you able to drive yourself like as a human without that sensor? If so, then the car should be able to do it. So I thought that intuition was really interesting. And so when I was thinking about this problem for tennis too, I was like. Okay. If I watch a match on tv, I can actually tell where the ball is landing.
I can actually, if I really wanted to, I could like manually chart this to some [00:13:00] degree of accuracy. And it probably would be good enough for an amateur to know okay, I had 50 serves that went in 20 serves that landed long, right? Like it, it seems like I should be able to do this just using my eyes and just using this one vantage point that they show on tv. And so I think I tried to take that same intuition, which is okay. Why don't I push this to the extreme and if I were able to get a lot of footage and train a model to detect these things that I can determine by just watching with my human eyes, could we like build a system that's like accurate enough for amateurs. And so that was like how the process started. And then I realized oh shoot, like I actually have been working on this problem already at Tesla. It's very similar. And so I actually think I could probably just do this myself rather than waiting for someone else to make this solution. And so then that's how. The idea for Ling vision came about and then like actually took the leap of faith, to go down that road.
Jason Jacobs: Got it. And what year was this when that chain of events started?
Swupnil Sahai: I'd say like around 2018, so about a year into working at Tesla, I felt like I had understood kind of the whole process that [00:14:00] was involved in training these kinds of models, and I was almost becoming more of a, I started as an engineer working in a very specific, solid problem, but then. As I worked at Tesla, I within a year, a lot happens I guess at these companies that move quickly. I was already moving to almost like a PM type of role where it's like I actually saw the full process and like what would be involved, what kind of team would I need to put together? So I think I started to have a zoom out understanding. And so I think around 20 towards the end of 2018 was like when I started thinking about okay, I think I could maybe put together a team to do this and I know what would be required, what resources are needed to achieve this.
Jason Jacobs: Dummy question, but there's a lot of talk about agents and. And what is an agent and everyone who's, many people who talk about it are, have a different definition in their head. And so there's a lot of debate about what agents can and can't do when people are, two shifts passing in the night because they're talk they're not operating off of the same definition.
When you say train a model, do you feel like that's a similar thing? What does it mean to train a model and doesn't mean different things to different people? And also has that definition changed as the big LMS have come into play over the last few years?
Swupnil Sahai: [00:15:00] Yeah, I think the definition hasn't changed too much. But when I think of training a model, it's really comes down to getting a bunch of data and having that data annotated. So it depends on the problem space, but like for us, it's like, it model needs to take in like pixels from it, from a. Image or from a video, and it needs to output like the location of the ball.
That's like fundamentally what it's doing. And I need to have lots and lots of examples of video footage of, shots going back and forth. And then the output of like where the ball actually was. Like that's essentially what to do is like curate that data set. And then if I can do that, I can teach a model.
I can train a like AI model, a machine learning model to taking that input and then. Put that as its output. And so if you look at language models today, it's actually similar in their case the input is like a prompt and then the output is a response. And then the output is done as like a, next, approximately, like a next word response. But it's a similar idea. And so it's have an input of text, what's predict what the next text should be. And so in order [00:16:00] to do that, you have to have a similar data set of input and response. And what's nice with something like a chat, GPT is we have so much data of this publicly available.
There's like Reddit, people ask questions on Reddit, people ask questions on Quora. There's and stack overflow, right? There's so many, like these publicly available resources where we have examples of like question and response or even just like long form text from an article that someone wrote.
All of that is actually training data that could be used to build such a model. So the training is actually the same, but in, in a lot of these LLM situations, there's a lot of data publicly available, which is really nice and it makes it easier, I would say. To be able to get that data for our situation.
There wasn't publicly available data of like video and like 3D data attached to it. Only the Hawkeyes of the world have that kind of data. No one else really has that. And so it's not something that we could just like scrape. So for us, we had to actually build that data, set ourselves and annotate it ourselves.
So it's a little bit extra work that we have to do. But ultimately it's the same process. And then I think with agents that's just like a new type of model [00:17:00] where. Your code can actually like, take action for you basically. And so that's like a whole nother mechanism I would say, like another layer on top of all this. But yeah, that's, I think training models is still, hasn't changed that much, but I think the architecture of the models has changed a lot. That's really what's made like LLMs so powerful and so good recently is like some great research and development that happened on the architecture of the model itself.
Jason Jacobs: And tactically, how did you go about building the data set in the early days?
Swupnil Sahai: Yeah, there's two parts to it. There's like the input, which is like the actual video footage, and then there's like the output, which is like actually like annotating, was the ball in or out, for example, just as like a basic level. And so I. Getting that footage. The first thing we did was we did scrape some footage from like YouTube and things like that because there was people uploading, like highlights or they'd upload their own videos.
They'd record their matches with a tripod and put it on YouTube for just their friends to see. So there's, there was actually a lot of public footage on YouTube. We just pulled some of that initially. But we still had to annotate it ourselves. So we had our own team of daily laborers that would like literally watch this video. And annotate every single thing that we wanted to model to output. So this is a [00:18:00] forehand, this is a backhand, this is top spin, this is slice. The ball landed in the service box on the other side on the dee side of the court. So all these specific things we had to annotate
Jason Jacobs: Would, it would it still be humans doing that today? As AI is starting to advance.
Swupnil Sahai: It depends on the problem. If it's something that's more that's been researched a lot and there's like lots of public data sets or public models available, you could automate some of it. So for example, like we have a player tracking model as well. I. Which focuses on tracking the players.
And so for a lot of that you can use some off the shelf like pedestrian detector to build the base for that. And then you have to add some additional things on top that you want to annotate specifically for your use case. So there are some like base models that you could use depending on the task. And for example, one of the things that to do is like label the ball and draw a box around it.
So that's probably like easy to do now with off the shelf models that you can just find online or open source. But the more specific your task goes. And the less publicly available data there is for that task, that's where you would need to actually label that yourself.
Jason Jacobs: So is there like a models marketplace? Like where does one go to find a model to meet their [00:19:00] task when they're determining whether to use something off the shelf or build their own?
Swupnil Sahai: Yeah, I think a lot of companies are putting out open source models like Facebook or I guess Meta has, most of their research and LM recently has been all like open source. Obviously we're seeing some companies in China doing that now too. There's I'd say private companies that will do research and publish their models as open source.
But then there's also a lot of just like academics that will be doing research and their publishing models. Typically the data sets won't be as large 'cause they may not have as much resources, but it may not need to be like super large for depending on the task. But yeah, sometimes it's just a matter of finding an open source model that was created through like some paper in a researcher institute. university, so it really varies.
Jason Jacobs: So you started by by scraping footage from YouTube and then doing the annotation yourselves, and then it sounds like there were some other aspects to to, to building up that initial data set as well.
Swupnil Sahai: Yeah. So in addition to that, we started going around the San Francisco Bay area and we would just like record people playing tennis basically. So we'd go to like academies where these really good junior players and we'd go to [00:20:00] just our friends matches. We started recording ourselves obviously. So I think it's just a variety of. Of data we started getting and so that was enough to build some initial version of the app that could do some basic stats Tracking. It wasn't like super accurate because it was so specific to the data that we had, and it wasn't like very generalized, I would say. The diversity wasn't great.
It was like only courts in the Bay Area, so it's it's very limited, but it was enough that we felt let's just put it out there and let's start getting people around the world recording. And then that's where the training data really started to grow. So then at that point, it's like anybody in the world can set up their phone.
I. On a tripod or on a fence mount, and they can record their game and that footage gets sent to us. And it doesn't matter where they're playing the lighting condition, the court surface the unique technique that they have, all of that gets put into our data set now. And so that suddenly, just blows up the volume of training that we have.
Obviously we still have to go through and annotate it. And at this point, what's nice is we built. Those like base models of our own that we can semi automatically annotate new data as it comes in. So we're just able to annotate faster and faster. Now [00:21:00] we basically built our own models to do this. But but that's how the tier data started to scale. So I think we took the Tesla approach, which was, they said, let's just stick cameras on all these cars. Let's have customers all around the world drive in all these different roads and we'll get all that footage back from them.
That was the approach we're using at Tesla. So we tried to take this same approach here, which is let's just get data from customers. That's probably the easiest way to do it. And then. It ensures you get really good diversity from all across the world.
Jason Jacobs: And you say we. So did you have a team initially and were you still at Tesla at this time?
Swupnil Sahai: No I wasn't at Tesla when we started, like the data collection part. I would say I was at Tesla when I was like starting to pitch to investors and that's when he got like Andy Roddick, who was like one of our first investors on board. We got some professional players and things like that.
But yeah, no, the team wasn't until I left Tesla. So initially my co-founder was my roommate from Berkeley, actually from our undergrad. So I've known him forever. So he's our CTO still today. So he's my co-founder. But then yeah, we started to build an initial team of people a couple ML engineers, like some interns that were like labeling the data, couple full stack engineers for building out like the app and everything.
Yeah, so pretty small team [00:22:00] initially, but that was like pretty shortly after. So this was like the summer of 2019, so almost six years ago. In June of that year.
Jason Jacobs: And so the initial app that you put out the door what was the MVP feature set?
Swupnil Sahai: Yeah, the MVP feature set was basically you would record your game and then at the end of the recording you'd get like a heat map of where all the shots landed. And so you could see a heat map of where they landed. You could already filter by show me just my forehands or just my back ends.
But that was the MVP was just like a stats readout. And then we just evolved it from there. So then. after that, we said, oh, it'd be fun if we showed this heat map on top of the video, as an overlay in real time. And so as you're playing, you can see like the shots being plotted, as you're hitting. And we thought that'd be fun. And then that became like super popular. Everybody really likes seeing the overlays because it made their video feel like more professional. And then we added like the speed of the shot on the video as well. And every shot you hit, you'd see a little speed pop in. And so people really like that. So I think it just made it more fun, I think. And, it wasn't necessarily like something that you're using that feature for, like improvement necessarily, but I think it's it just [00:23:00] made your video feel more like a professional match and it's oh, I'm getting the same stats as the pros.
I can see my serve speed now. So I think, as we just listened to customer feedback, we started adding more things like that, and then eventually we just started adding like highlights and we could remove all the dead time in between the points. When you're like. Picking up the ball or getting a water break, we can cut down the match to 20 minutes now, a one hour match, so all of that we started adding just based on customer feedback.
But but yeah, the initial version was basically just like a stats readout.
Jason Jacobs: With the benefit of hindsight, do you think like scrapping together an initial data set and putting something out the door and then turning the users into the data collectors? I mean should every I mean is that the way that every data product should launch or how do you know if you are innovating in an area that could benefit from this launch strategy and where might not it make as much sense?
Swupnil Sahai: I think any AI technology sh, needs to take this approach. And I mean if you look at it, even Chad, GPT and these big. These big companies are all doing that too. Like they usually have some [00:24:00] feedback like, was this, was this helpful? Was this not? So all of that is input that's helping them improve the model and like fine tune it and everything.
Yeah, so I think like any, a company should try to get a product out there. I think the bar is getting higher, it feels like. As these models get better and better, people have like less patience for the inaccuracies or the hallucination that they see from these models. So I think that's, that might make it a little bit tougher, but I think what one advantage you have if you're going in a niche is like, there's not a lot of other players and so there's nothing else that they can really compare it to, there's no competition.
And so then you, the bar's a little bit lower than in that case. you want it to be like somewhat accurate initially and provide some value, like that's the most important thing. Am I actually providing like a net positive value to this customer? I'm not just wasting their time here with this product.
I think if you're, if you meet that bar, which is a very low bar, I feel like that's enough for an MVP and like just get it out there. Start getting more customer data, get more customer feedback, and then just iterate from there. And it might take a while. Like it took us probably like at least. A year to a year and a half of having launched a product before it actually started to get [00:25:00] somewhat reliable. And then now, in terms of being able to call the lines, that was probably like maybe three years after launch that it actually got accurate enough to call the lines. So yeah, all of that kind of builds up over time. But I think I'm always in favor of kind of shipping sooner rather than later. and just getting feedback. But I think especially for any AI company it's really helpful to just get like actual customer usage
Jason Jacobs: And then what? And then what does that iteration cycle look like in terms of. How how do you get more accurate over time? So I get as there's a bigger data set, you have more to pull from, but then what do you do with that data that enables you to continue to drive efficiencies and make it better?
Swupnil Sahai: Yeah. So usual what will happen is once you get some, critical mass of data, then there's this long tail of just like edge cases where the model struggles. And so that's where like user feedback is really helpful. So for example, within SwingVision, we added the ability to actually edit the shot data that was coming in.
So for example, let's say you had a rally and then there was some shot that was like misclassified as like a backhand when it was actually a forehand. You could actually [00:26:00] go in and edit that shot and change that. And so we added that kind of ability. And then our system, our backend, would flag it anytime somebody made an edit.
We knew that, that portion of footage was like interesting 'cause the model was struggling with it. And so it helped us source more data from those examples. And I think Tesla's done the same thing with autopilot. So it's if you're using autopilot. And you turn off autopilot, like you hit the brakes or you like swerve out of the way you're having an intervention.
And so that's a, it's an interesting piece of footage for them because the model didn't behave properly and human had to take over. So I think that sort of corrective feedback is really helpful. And then that helps you focus your data. 'cause at a certain point you have like too much data and you don't necessarily train on everything.
'cause like it's diminishing returns at a certain point. But those examples where you know you're struggling and your customers are literally telling you that the model is not performing well. That's a really good set of data to include and train more on. And you'll see this with any sort of chat interface as well.
There's, you have opportunities to provide feedback and, what was wrong or what could be improved, and that's all really valuable for these companies.
Jason Jacobs: And what was, it sounds like you raised money r right from the [00:27:00] outset. What was the logic bet behind raising money? How did you think about that raise, like from who and what it would be for, and then how has your thinking on. How you capitalize the company evolved over the several years now that you've been building.
Swupnil Sahai: Yeah, I mean it was interesting, like we actually tried to go straight out for venture capital initially. But most venture investors thought either this was just such a new category that they weren't sure like what the market would really look like, market size, and the technology also at the time just seemed like really hard to develop.
Are you really gonna be able to track all this with just one camera and just the phone it doesn't really seem possible. There was no like analogy of oh, someone's already done this in like or whatever. And it's we're just doing this for like racket sports. So it was really tough 'cause I think we were like creating a new category. And so that basically crossed out an entire market of investors for us. And so then we just focused on friends and family and like angel investors like professional. Former professional players and people that we knew just through our network. [00:28:00] So yeah, the initial round was just like two 50 k, really small, and it was just enough to get like a few, like interns basically, and like part-time people helping us out. And then once we shipped the product, we raised like a slightly larger angel round. And then we got like some more full-time people on board. And then there were a few more angel rounds, since then. But yeah, I think in every round we tried to just really minimize, like dilution, minimize how much we're raising, like just keep the team like super lean. As lean as possible. so it was like a blessing. I think that we didn't go through the venture path initially because it also did take time for us to get to product market fit and like for the product to get like accurate enough to get there. And so I think like raising a lot initially for us at least, maybe would've been, worse for the trajectory of the company. And then we did end up raising venture eventually which was two years ago. So that was like our first like formal like price round led by a VC based in the Bay Area. So like that was the first time that we actually, started to take on like more significant capital, I would say. but up until that point it was mostly angel funded.
Jason Jacobs: Uhhuh and and how [00:29:00] have you. Gone to market in terms of how did you get the initial customers and then how do new people find out about your service today?
Swupnil Sahai: Yeah, initially I would say the App Store and like honestly, like Apple was our biggest distribution partner because. We were, I always exclusive and some people at Apple like noticed pretty quickly when we launched the app that it was just so innovative what we were doing and we were taking advantage of all of the hardware, the neural engine, everything.
Like we're really pushing the limits of what you can do with your phone. And so they noticed that actually pretty quickly and I'd say within a year of launch they started featuring us in various places. So first we started getting featured like in the app store as like app of the day and things like that.
Then, 2021. So basically like two years after launch, we got featured in like multiple keynote events. So we were like an iPad keynote, iPhone keynote. So they started showing us, like showing SwingVision demos together in the launch of the new iPhone, which is like crazy. So all of that gave us so much credibility and exposure that like we didn't really spend anything on marketing for the [00:30:00] first couple years.
'cause we didn't really have to. And it was just like all word of mouth. And then at some point we had started to build this like following on Instagram. because we were sharing like highlights from players on Instagram, and so our Instagram page just organically started to get like larger and larger.
And so we realized like maybe paid ads on Instagram could be like a channel for us. And so then that eventually became like our largest paid channel for growth. But even today almost six years later, like it's still primarily organic. It's still primarily word of mouth. It's probably 75%, honestly, organic still. So it's either people will record with their friend and then they'll share the video with the person they just hit with, right? So this is like natural virality to the product which has been really nice for us. And then on the other side, it's really common that people will just post their highlights on youTube or Reddit, and it's really common if you go to there's this r slash tennis, like the number 10 s, which is like the tennis sub Reddit on. And if you go there, there's so many videos that are posted about SwingVision all the time. And it's not really like about SwingVision, it's just oh, here's my serve. What do you guys think? What can I do to improve my serve? So I was just trying to [00:31:00] get like community feedback, which is pretty cool, but it's pretty much always a SwingVision link. And then on YouTube people will share like longer match highlights and things like that, and coaches have started using it too.
So those have been some, like some other sort of ancillary sort of channels for us too. But yeah, most of it's just word of mouth and organic and, ideally we keep it that way, but we're always looking for, new pay channels to accelerate and add more fuel to the growth.
Jason Jacobs: Huh And we talked a little bit before we started recording, but you guys got going bef before there were these big LLMs. And before, AI as we know it today, had really taken off. What are the things that if you were starting the company today, you would do. Different. Given how the landscape has shifted under your feet and also what, if anything, are you starting to do different now as these tools are continuing to get more powerful?
Swupnil Sahai: Yeah, it's interesting. I think there's been, I guess just at a more technical level, like a really big. [00:32:00] Architecture change in these models. The ones that are being used today for these like image generation tasks even for just other vision tasks the state of the art is actually a very different architecture from what we're using at SwingVision. And part of that, part of the reason why we can't quite use that is because we're doing all our processing on device. So it's really important that the models are like lean and lightweight and really fast and efficient, but this breakthrough that's happened recently with these models it's impressive in terms of the accuracy, but it is very expensive.
Like all of these LLMs are massive models. They require lots of resources to run. Most of them it's, very difficult to have that level of accuracy and squeeze that down to run on a phone. Pretty much all of them are running in the cloud on some like giant GPU server. The landscape's changed a little bit. I think if we didn't know what we know, we'd probably be starting today in those new architectures. but that may not actually be the best solution depending on the problem. I. For most, I'd say 90% of use cases, it's probably fine to process in the cloud because you don't necessarily need such quick, results right away.
And the data [00:33:00] that you're sending, especially if it's like a text-based product or service that text that's being sent to the backend can be sent very quickly. So like the latency isn't a concern there. But I think if it's a product like ours where we're like recording video live, like sending those pixels out in real time would be really difficult. And so that's where like on device processing is really important. So I think. We're in a sort of different category where we're focused on-device processing, and so therefore, like it constrains the kinds of architectures you can use and the kind of compute that you have access to. Which has, I think, been interesting.
But we're always evaluating, what's come out. We're always looking at the latest research papers and seeing does it make sense to, to shift things or try new areas. And yeah, and then we're obviously trying to integrate AI more into our workflow in general, just as a company. So I think that's like what every company is doing at this point, right? And so a lot of these tools are helping our engineers or even just non-engineers, like move faster. And we're always a little bit careful about it and especially 'cause we just have so much like proprietary information and data and everything, and so we're always like really cognizant of that.
But I think in general, like [00:34:00] it's becoming a necessity to be able to leverage these new tools. Regardless of what role you have in the company, so we're, we are trying to encourage that, but be, a bit practical about the like kind of safety and IP concerns as well.
Jason Jacobs: I've, I've of course heard that from other people as well, especially the people that are working like. In healthcare, for example, where you're, you've got patient data and things like that. As a potential founder who's evaluating building companies, using these technologies really in any space, what advice do you have just in terms of how to think about privacy, what precautions to take, and how to be eyes wide open without, writing things off unnecessarily or shooting yourself in the foot.
Swupnil Sahai: Yeah, I don't know if I have the best answer for that question, but I think the policy that we've at least we're trying right now with our engineers is like we're avoiding any AI tools that expose like our entire code base. I. But at the same time, it's okay to paste in small snippets.
If you're trying to get a specific thing answered or, improve some specific part of the code base or solve some very specific problem, I would say, that's [00:35:00] okay. But I think we're just really hesitant to have any tools that are like just full on exposing like our entire code base. I think that's just where things can start to get problematic and there's probably like some middle area in between that's a bit more reasonable than that. And I think more of these more of these companies are coming up with like local approaches where you're, you could have a model that just runs locally on your computer, and so like your code base is not necessarily being exposed outside of your local, environment. And I, hopefully more companies will go that way. So I think people are more comfortable. but it's it's like anything, technically everybody's code is hosted on GitHub or Bitbucket or whatever, so it is like it is out there, right? So it's like you just have to evaluate each technology, really carefully. And take a look at those like service agreements and the finer details. Are they allowed to train on your data? All that's really important because if they are able to train on your data, then that's makes it a lot easier for like a competitor to, to be able to replicate what you're doing.
So I think it's, that's where it can get a little tricky, yeah, I think being cautious, but at the same time you don't want to just fall behind. So I think there's a delicate balance there. 'cause everyone's using these tools to some [00:36:00] extent and some more than others.
Jason Jacobs: When it comes to business model, so is it a freemium model today?
Swupnil Sahai: Yeah. It's a freemium model we've really modeled it after a Strava and all trails. So it's you can download the app for free. There's some decent amount you can do for free. I'd say we probably are leaning a bit more on the monetization side. We always have, because we weren't like, we didn't have a ton of funding initially.
We're almost like bootstrap slash anal funded, so I think we always erred more on the side of monetization. And so because of that, yeah. It's like you can record two hours a month for free, which is basically just giving you like a taste of what the product can do. But if you wanna record more than that, if you wanna have cloud storage for all of your videos, access more advanced features, like the line calling then we have this pro subscription that people subscribe to.
And yeah, in our case most people subscribe annually which has been good for us just from a cash perspective, obviously. That's really helpful. But yeah that's pretty much how it works is just a subscription product.
Jason Jacobs: And when we chatted before the recording, you talked a little bit about the the experiment with the [00:37:00] cameras actually at the facilities versus using your phone. I'd love to hear more about that experiment and also any other experiments that you either have run are running or may run in the future to the extent that you can share.
Swupnil Sahai: Yeah, for sure. I. Yeah, since the beginning we've been direct to consumer. That's, how we've been able to have such great scale of our training data and just general usage. The fact that anybody can just use their phone, it's excellent. But there is a little bit of friction to setting up your camera, especially as a first time user.
Most people will get over that hump after they do it a couple times. But we're always just looking for ways to, like, how can we improve every part of the funnel, right? And so it's like, how can we reduce friction? How can we make it easier for you to set up? And we've always added things over time. In the beginning we would just suggest like some tripod that you can buy on Amazon or some fence mounting device that you could buy on Amazon. But eventually we actually found one solution that was working really well and had really high retention. We actually saw the retention curves based on which equipment it used, and we ended up acquiring the patent for that design for that solution. I actually have it here [00:38:00] on, on video. I don't know if everyone's gonna be watching this or listening, but. It looks like this. So it's kinda like a selfie stick, but it has a little hook here. And so your phone sticks in here just a selfie stick would. But what's interesting is this part slides into the top of the fence.
And so this is like this telescopic pole that you can extend and you can just hang it behind you and hook it on the fence behind you. And so it's a very kind of standardized way to set up on any tennis court and I guess pickleball court now too. and so we ended up acquiring the patent for the design.
We made our own version of it that's like sleek and like all black and has our branding, all this stuff. So we're always finding ways to improve. Even this one, actually today we're trying to update it. So currently you have to like the most friction is like figuring out what tilt your camera needs to be in order to see the court. And, you might stick it up on the fence and then you get some audio feedback like, oh, it's not angled properly. Tilt it down more
It back down, adjust it a little bit. Try it again. So it's just this back and forth you have to do sometimes. And you're making a new version that has a motor inside, and so
Jason Jacobs: eh.
Swupnil Sahai: you just put it up and it's gonna automatically like a little robot, [00:39:00] find the court for you.
Jason Jacobs: That's cool.
Swupnil Sahai: always thinking of ways like this to just reduce the friction. I think yeah, thinking longer term, what is like the t most frictionless experience for SwingVision would be that like I just walk on the court and there's already camera there. I don't need to set up anything.
I just like tap start on my phone or like a kiosk or something and the recording just goes. And so that's the most frictionless experience. Obviously that's a little bit more complicated for us to provide because it involves some infrastructure being there at the club and, this is maybe hard to see happening at every facility today, but.
Maybe in five, 10 years you could imagine like every court has this has a camera on it. And so we're trying to move towards that vision now. And so that's like a new thing that we're trying with some clubs installing fixed cameras and essentially making it frictionless. And, we'll see how that develops.
It's a, it's definitely a different customer acquisition model. There's also like a sales motion attached to it as well to convince the club and get buy-in from people and all that. So it's obviously much slower, but. It does feel like long term that is like the best possible experience you could have is that I don't [00:40:00] have to set up anything.
It's just every court has a camera that's just how it should be. And so I think we're hoping to get, start going down the path of that vision, but unclear how long that would take to, to achieve. But that's the long term goal
Jason Jacobs: I don't know if this is a good idea or a bad idea, but one thing that comes to mind is Peloton has their like flagship studio where anyone can go and visit. It becomes like a tourist destination. Maybe you guys like, start by controlling everything and build your own facility and it becomes, the hubs will run all your experiments and stuff, and then you can use it as like the demo facility to get other facilities on board.
Swupnil Sahai: that's really cool. That's an interesting idea. Yeah. I'd love to have some sort of like retail presence someday. That'd be amazing. That's a
Jason Jacobs: And I've been thinking a little bit about that in the hockey world too. It's a lot of these training facilities for skills, they don't have cameras anywhere and they're really, it's all kind of humans that are, there's no like stats or anything that you're talking about.
And, and imagine like either building or buying facilities and then driving efficiencies and making them really tech forward. And then not only driving out a bunch of the cost, but also improving the experience and then adding a whole layer of [00:41:00] intelligence that doesn't exist.
Today. So that whole theme of like private equity meets SMB with roll-ups and stuff like, could you do that with tennis courts? Could you do that with with sports facilities of different kinds? It's not, I don't know. It's not my world, but but it's just an interesting thread to pull in.
Swupnil Sahai: No, definitely. Yeah. Yeah. No, it's cool. That's it's certainly an option, I'd say so. Yeah, that'd be a nice place to go to.
Jason Jacobs: Got it. And what does the future look like for you? So if you think about, two years, three years, five years, 10 years, are you still racket sport focused? Are you still a hundred percent or largely software? Are you still subscription model? Are you doing anything else? And if so, what might that be?
What, I guess said another way what does your ambition look like in its fullest form?
Swupnil Sahai: Yeah, I think. Probably the midterm, two, three years. It's probably still rocket sports, but ideally more sports. So we have tennis and pickleball today. Paddle is fastest growing sport outside of the US right now, so it's like all over western Europe [00:42:00] and Latin America.
It's like this kinda like smaller tennis court with glass walls around it. So it's like a hybrid of tennis and squash. And so that's gonna be our next sport we're launching this year actually. And then from there we'll see, what makes sense. I think we. Definitely like to continue expanding to more sports.
There's so many underserved athletes and all these other sports. I think the most natural ones after that would probably be other racket sports, like table tennis or badminton. But those aren't quite as big in the us. They're more, bigger in other parts of the world. So it depends a bit on like how our, just general usage and organic uptick is in these other countries. Which it is a global app already, but I think still primarily like US and Western Europe, I would say today. but yeah, I it's hard to say exactly, but I could see us, going beyond racket sports to other ball sports like baseball or volleyball potentially even team sports.
But I think that's a, maybe a more of a stretch from art from where we are focused right now. but I think what we really like to get to a place where is like any sport that we're in, we're really like. Up and down throughout the entire vertical. And so it's [00:43:00] like when you start playing, this is like the first time you ever play you would be recorded with SwingVision and like the rest of the time that you're having your experience with this sport.
Whether it's like taking a lesson playing in a competitive game and like high school, playing in college, playing as an adult with your friends. Like no matter what sort of scenario you have when you're playing this sport, like it's being recorded with SwingVision, it's being shared, it's being streamed, potentially even, that's like the vision for us. And I could see eventually, like we could even go farther up market in the sport too. We could eventually build swing engine for like higher levels of the sport and maybe even get into the pro space someday. And I think that would be such a cool, story to that. It's like the same technology that you're using as what pros are using. I think we could get there, we'd probably need more cameras to do it. But I think we could do it for cheaper than the current solutions that are out there, especially just with the lead that we have in our training data ai.
And so I feel like there's a potentially opportunity for us if we want to go down that path to just be pervasive across each sport that we go into. [00:44:00] And that would be, I think, an ideal outcome for us.
Jason Jacobs: So it seems like directionally the cost of building consumer apps or, software in general is continuing to fall. The speed at which you can spin up new applications is continuing to compress the size of the team. You need the capital requirements, like everything is getting smaller.
How do you think about durability in the long term? Because durability, it's like you need to grow your customer base. You need to have happy customers, you need to keep those customers as new products come along. And so in a world where there's more and more that are better and better that you're competing with like how how do you think about.
Defensibility and how do you think about durability and how do they interrelate? Do you need to think about defensibility to have durability, or can you just build a great product and get all the details right? Is that enough?
Swupnil Sahai: Yeah. That's so interesting. I think for us right now, [00:45:00] it's the big defensibility for us is probably the training data. That we've built up ourselves. And I think that's gonna be really hard for any company to replicate. I think there's not like publicly available data necessarily, or like foundation models that would necessarily get you to where we are or nowhere near the accuracy that we have, I would say at least. So I think that's like a moat that just keeps growing for us as we just get more and more
Jason Jacobs: That, that reminds me of the MyFitnessPal guys in the early days where their remote came from with the food tracking, they had all the foods.
Yeah.
Swupnil Sahai: Yeah, exactly. So it's someone could theoretically build that, but it's like you would need to move like really quickly and like somehow get to have more, acquire more customers than we do and have more footage than we do and label it faster than we do.
Like you would just need a lot of resources. Obviously it's possible, but it's like hard and I think it gets harder and harder over time. But I think that's a big defensively for us. And I think like longer term, it's probably just more like our brand and like the trust that we have. From customers that like, they know they're gonna get a reliable, accurate, dependable experience with our brand.
And so I think like longer term it's probably the brand that you [00:46:00] wanna stick with and brand loyalty that you want to have. And that's what like companies like Apple obviously have done so well with. So I think, then it's a matter of, okay, when you have lots of different players, it's probably like brand that's gonna keep you ahead. You look at like a Nike, right? These kinds of companies like this just. There's hundreds of those kinds of companies, but like only a few are always there and durable. But I guess brands do move up and down and new ones come up. So it is interesting.
Jason Jacobs: Yeah, but I think what I'm hearing is that look, you can give leverage to the process and stuff, but they end. But at the end of the day, the fundamentals are the fundamentals, and that's not gonna change. I think that's what I'm hearing.
Swupnil Sahai: Yeah, I think so. And I think there's also something to be said about just like having a better user experience and like having a particular taste in the way that you build your product and the way that you design it. So I think like people do care about that, and I think it does
Jason Jacobs: I.
Swupnil Sahai: you some meaning to your brand as well.
And I think that's something we've also always been like really focused on too. So I think, and that's something that we've heard from customers that like, even as new players have come up in our space, no one's come up in like the mobile phone space. There are other like fixed camera solutions already there [00:47:00] today in all these sports. But what we always hear is just it's so clear that like we just have like really good taste in our ux and like the way that we respond to customer feedback and build it intuitive, experience. I think that's also important. And as I guess later on it's more important than early on.
I think all those things will compound over time as well.
Jason Jacobs: Is there anything broken or missing that is outside of the purview of what you might address, but that if someone else did, would make you guys go faster and more effectively?
Swupnil Sahai: Oh, that's interesting. The one thing that comes to mind is just the prevalence of this incredible silicon that Apple has made. I would say one reason we're not on Android is because the silicon is just not as efficient, not as powerful on those devices. So I think I'd like to see more mobile hardware get cheaper and get. Get more powerful and still stay efficient for mobile devices. That would certainly help us. And I think other companies that are trying to do this kind of on device machine learning which, not, [00:48:00] is not, like I mentioned before, it's not necessarily the case for every company, but I think certainly would be for us. but yeah, in the sports space in general, it's hard to say, but I think there's just, there's lot, there's lots of opportunity and lots of like low hanging fruit. I feel like. Like recruiting is still done the way that it's done today. Coaching is still done the way it's done today, and I could see various new technology solutions that would improve all of those processes. And then would just, lift everyone up, including us. And that's not necessarily be competitive with us. So I think there's a big opportunity there. Like training, the way that people train today, especially like in a sport like tennis, you just go out and hit serves like, it's so boring.
It's just how could we make that more interesting? So I just feel like there's. a really big opportunity for new businesses to come up that would still lift us up too. And that wouldn't be competitive. And so it's an exciting time and I think as you said, because it's so much more accessible and like more affordable to be able to start these things on your own and maybe even bootstrap your own company now, like it's really cool to see.
I'm excited to see like what new ideas are gonna come about and what solutions will come to market. But it's a fun time to build, I would say.[00:49:00]
Jason Jacobs: I agree. This was a really great discussion. Is there anything I didn't ask that you wish I did? Or any parting words for listeners?
Swupnil Sahai: No, just, yeah. Thanks. Thanks for having me on, and I'm excited to see what you end up doing as well. And I hope more people will give it a shot and, es especially for someone like me who again was always very risk averse and maybe more introverted. If you have the ability and I guess like the privilege to be able to take a risk, I think I'd just always encourage more entrepreneurship in general. It's it's not for the faint of heart, but I think it's rewarding ultimately. And I think as you said, like the barrier's never been lower to be able to enter and become an entrepreneur. So this is the time. And if you have happen to have the position where you can take that risk take it and just give it a shot.
And there's just so many problems that need to be solved in so many different domains, not just sports. And I hope people listening will like, give it a shot and just try something new.
Jason Jacobs: That's a great point to end on. And for what it's worth I wouldn't have predicted from this discussion that you were [00:50:00] risk averse or an introvert. Yeah, you've,
Swupnil Sahai: me to be less of an introvert, but I think some things you can't really change yet at heart. But yeah, I'm trying
Jason Jacobs: I.
Swupnil Sahai: a more extroverted.
Jason Jacobs: Well, Swupnil, thanks a lot for coming on and sharing your learnings. I'm excited to check out SwingVision as a user. We, my family talks about playing more racket sports. We've not quite we've not quite broken the ice yet, but maybe this summer, we'll be this summer. But yeah congrats on your success and looking forward to following your progress.
Swupnil Sahai: Yeah. Thank you so much again for having me. Super fun.
Jason Jacobs: Thank you for tuning into The Next Next. If you enjoyed it, you can subscribe from your favorite podcast player in addition to the podcast. Which typically publishes weekly. There's also a weekly newsletter on Substack at the next next.substack.com. That's essentially for weekly accountability of the ground I'm covering, areas I'm tackling next, and where I could use some help as well.
And it's a great area to foster discussion and dialogue around the topics that we cover [00:51:00] on the show. Thanks for tuning in. See you next week.