David AI: Powering the Voice Era of AI

1. 0:00 – 0:12

   Introduction

   1. DHDiana Hu0:00

      Welcome, everyone. I'm excited today to have David AI here, who went through the batch in summer '24 and just announced their Series A for $25 million.

2. 0:12 – 0:31

   What is David AI?

   1. DHDiana Hu0:12

      So tell us, uh, Ben and Tomer, what David AI is.

   2. TCTomer Cohen0:16

      Thank you. Yeah. Um, yeah, David AI is an audio data research company. Within audio, we're focused on speech, and within speech, uh, conversational data, so conversations between people talking in different languages, dialects, accents, different contexts of

3. 0:31 – 1:11

   Challenges in Audio Data

   1. TCTomer Cohen0:31

      conversations.

   2. DHDiana Hu0:31

      Why is that so hard to find? I mean, it works for LLM where from text from the internet.

   3. BWBen Wiley0:36

      We found that there's no real, like, common crawl for audio. And on top of that, most of the audio that's on the internet is mono channel, single track, uh, which is not the format that the kind of bleeding edge model architectures for end-to-end speech models need and demand. We went really deep in trying to understand what solutions there are, uh, off the shelf for being able to separate audio that isn't separated at the source, and it just wasn't sufficient. These models have very, very low tolerance for any sort of bleed between channels, and we identified the only way to get high quality data was to collect it separated at the source.

   4. DHDiana Hu1:08

      It's a little bit interesting how this company ended up

4. 1:11 – 1:46

   Origin Story of David AI

   1. DHDiana Hu1:11

      getting started. Tell us a bit about the origin story.

   2. TCTomer Cohen1:13

      Ben and I met at Scale and, um, became close friends there and kind of first and foremost were excited to do something together. And then we also started getting really excited about multimodal AI and specifically voice AI as kind of like, uh, the next evolution of bringing AI into the real world. Um, so we applied to Y Combinator while, while we were still at Scale, uh, with just an idea and, like, something to explore, and we got in, so thank you for that. We left our jobs and then moved to SF and immediately started

5. 1:46 – 4:12

   Building the First Product

   1. TCTomer Cohen1:46

      building. The way that David AI came to be is we started off by reaching out to lots of YC companies that were training different kinds of multimodal models to figure out what kind of support they needed. And there was one company in particular that was training humanoid, humanoid robots, and they were really excited to talk to us, and the thing they needed most help with was, um, audio data for the robot's voice. And for us, that was a bit of an aha moment that this, like, robotics company that's solving all these really hard kind of physical world problems needed the most help with audio data.

   2. DHDiana Hu2:22

      By narrowing down and just focusing on audio and voice, it ended up secretly being a really good idea. It's a bit contrarian because people would think, oh, Scale has the whole market. It is done.

   3. TCTomer Cohen2:36

      Mm-hmm.

   4. DHDiana Hu2:36

      What gave you the confidence? Because I think when I... when you guys were in the batch, you weren't sure. R-remember doing office hours, you're like, "Oh, this is staying with this, like, small startup."

   5. TCTomer Cohen2:45

      Mm-hmm.

   6. DHDiana Hu2:45

      And then you decided to go down the rabbit hole.

   7. TCTomer Cohen2:48

      Um, audio isn't just kind of calling customer support agents. It's also the way that you interact with robots in wearable devices, in games, in, like, avatars. Any kind of, um, real world AI use case where you're not interfacing through a laptop and a keyboard, like, requires some for-form of voice or audio. Um, so it's bigger than you'd think. And then the other thing is, I think people have thought for a while that to build a big company in the data for AI space, you need to kind of like horizontally integrate and just, like, cover as much data surface area as possible, um, and just, like, build a company that's very nimble and can, like, hop from one thing to another when the kind of the tides shift. But, like, we believe that the best way to build this kind of company is to pick a vertical and go really, really deep and build a deep product, solve the hardest kind of hairiest problems within that modality, um, and find repeatability that way.

   8. DHDiana Hu3:42

      You built this just in a weekend, right? The-- In terms of tech. What, what was it?

   9. BWBen Wiley3:46

      Yeah. We ended up building this phone calling application in-- over the weekend, basically to get all of our friends and family to call in and, and have some conversations to test out some of these kind of hypotheses of how to collect high quality audio data. And by the end of that weekend, we had our, like, first kind of small data set to go out and, and bring to the world. But, um, that has now since evolved to this massive worldwide platform where people are, um, conversing in scripted and unscripted conversational settings. So it's been pretty cool to see the

6. 4:12 – 5:24

   Early Success and Growth

   1. BWBen Wiley4:12

      progression.

   2. DHDiana Hu4:13

      You actually closed some sizable contracts during the batch. I mean, you started small, and then you got some big ones. Tell us about those.

   3. TCTomer Cohen4:19

      We started off, uh, talking to a few YC companies, and this robotics company was our, our first customer. It was a thousand dollar contract, so really small, but we were excited about it at the time. Through them, we kind of figured some stuff out and built a perspective on, like, audio and audio data, and we could bring that perspective to the next customer and then the customer after that. And by the end of the batch, we closed our first six-figure contract with a big AI lab, and things have kind of continued to, to build up from there.

   4. DHDiana Hu4:46

      And by continue, you mean right after, couple months later, you closed actually seven-figure contracts really quickly.

   5. TCTomer Cohen4:53

      Today, we work with most of the, the big tech companies too, and they have, you know, massive audio data needs. Um, and what's been exciting is to see how kind of our sales motion has kind of built on itself. And, um, kind of the more data we collect, the easier it becomes. I mean, really, like, we talk about this a lot internally. We never feel like we're selling. We kind of-- We have this data, and it's up to a lab to decide whether or not it's useful. Um, and if it is, then great. If not, then, then

7. 5:24 – 7:40

   Business Model and Approach

   1. TCTomer Cohen5:24

      no.

   2. DHDiana Hu5:24

      I think the cool thing about it is actually with setting up your company with focusing using audio and voice, you are actually building pretty much a audio data research lab. That's kind of how the company ended up being set up.

   3. TCTomer Cohen5:40

      Mm-hmm.

   4. DHDiana Hu5:41

      It was not something that you could necessarily do at scale, right?

   5. TCTomer Cohen5:44

      We think about ourselves as, like you said, a, a kind of a data research lab. Um, and what that means is we try to build our own perspective on where we think models should go and what shapes of data would get them there. Um, then we do internal R&D to figure out-if that data is working and once we feel like we've struck gold, then we'll like scale it up to a really big data set, and then we put it out in the world, and if we did our jobs right, then the model companies will adopt that data. Um, that's different than the way that most companies in the space operate, which is more of like a professional services kind of model, where a lab will go to them with a very custom kind of bespoke request. The data labeling company will kind of do a bunch of work to find the people and collect the data, and then deliver it over to the customer. The customer will own that data, um, and the company will take a, a take rate. I think that's how the space has traditionally operated, and I think that business model works clearly. Um-

   6. DHDiana Hu6:44

      And Scale made lots of money from it too.

   7. TCTomer Cohen6:46

      Scale's doing fantastic, and they will continue to. But we were excited to explore like a different way to go about it, um, like is there another business model and another operating model that works? And, um, we think we found one, so...

   8. DHDiana Hu6:58

      I think the really cool thing is that one of the fastest growing categories for a lot of YC companies that are taking off is AI voice agents for vertical businesses.

   9. TCTomer Cohen7:07

      Mm-hmm.

   10. DHDiana Hu7:07

       And at the end, it's all the way turtles down, and you're the reason why a lot of these models are doing well.

   11. TCTomer Cohen7:14

       Voice AI apps are only as good as the models underneath them, and the models are only as good as the data underneath them. And, you know, I think the, the data layer in particular in audio doesn't get as much attention because of all these like awesome voice AI apps that are taking off, and we've been excited to kind of build a bit under the radar, like kinda do the picks and shovels stuff, um, that's kinda propping up these, these applications in the space.

8. 7:40 – 8:50

   Future Plans and Hiring

   1. DHDiana Hu7:40

      So now that you're off to the races with all these big contracts, growing, what's next for David AI in the next five years, let's say?

   2. TCTomer Cohen7:49

      The most important thing to being an audio data research company is having a strong audio research function, so that's a big focus of ours in the next few months, um, and years is to, to build that out. Basically build the ability to, um, predict the future a bit or, um, know where we want models to go and what kinds of data we should go collect. And in parallel, like building the product and like the operation to scale up our data collection 10X and then 10X again. We feel like there's this massive opportunity ahead, um, for us to go grab, and our big focus right now is, is growing the team, um, across a bunch of different functions to enable us to do that.

   3. DHDiana Hu8:29

      You guys are growing a lot. You're hiring. Tell us more about the roles you're hiring for.

   4. TCTomer Cohen8:33

      So we're looking for, uh, researchers to kinda help us realize this data research company vision, uh, and predict kind of the future roadmap for data we wanna go and collect. Uh, and on top of that, engineers and operators to kinda help realize that vision underneath everything too.

   5. DHDiana Hu8:45

      Congrats on the Series A, and thanks for joining us.

   6. TCTomer Cohen8:48

      Thanks, Diana.

   7. SPSpeaker8:49

      Thanks for having us.

Episode duration: 9:00