Pablo Palafox, Co-Founder and CEO at HappyRobot | AI Minds #051

Welcome back to the AI Minds podcast. This is a podcast where we explore the companies of tomorrow being built AI first. I'm your host, Demetrios, and this episode, like every other episode, is brought to you by Deepgram, the number one speech-to-text and text-to-speech API on the internet today, trusted by The world's top enterprises, conversational AI leaders, and startups. Some of them you may have heard of, like Spotify, Twilio, NASA, and Citibank. This episode is a very special episode because we are joined by Pablo, the CEO and co-founder of Happy Robot. How you doing, dude?

Doing great, man. Thanks for having me.

Well, I know that you are from where I— the country that I used to live in. You grew up in Spain and then you migrated over to the country that I now live in Germany. You spent some time in Munich. What were you doing there? Yes.

So in Munich, I was actually doing my master's. I went there 2017 and then ended up doing my PhD there with Angela Dai and Matthias Niessner, 2 great minds on the computer vision realm. So that's how I ended up there. And I spent there like 5 years, man. It was a long time.

So the German culture, I like to make jokes that most people do the opposite of what we did. They go from Germany to Spain, but we went from Spain to Germany. And that is a little bit of a pattern interrupt. You glossed over something real fast that I want to mention. Your teachers, they just had a bit of success, right?

Yeah. Matthias Niesner is one of the co-founders of Synthesia, and I think they just raised a A little round, just started.

Massive.

$180 million. Nothing, nothing fancy.

What are they going to do with all that money? That is amazing. So congrats to them. And that's cool that you got to hang out and learn from Matthias. And then you decided to work in big tech for a little bit?

Yeah, during the PhD, I actually did a bit of an internship with Meta. It was COVID time, so it was remote. It was interesting. I should have come to South Salido actually. I was working with the virtual humans team, great folks. And I was doing like some, basically my PhD was around virtual avatars, if you will, 3D reconstruction. And yeah, I should have come here to South Salido, but I ended up doing an internship from a little village next to Munich. So that was interesting.

Yeah, that I can imagine. That's like me living in this little village next to Frankfurt, and I joke that there are more cows in my village than humans, and most people don't know what AI is here. That's, uh, so that's fair.

That's probably, I can, I can, I can see that. Yeah.

Yeah. It keeps your feet on the ground, that's for sure.

It is quiet. I mean, you, it's just like a different lifestyle. I mean, I remember When we actually, when I actually dropped out of the PhD and start thinking about ideas with my 2 co-founders now, we were in this little village and it was like great on one side. Like it's like you can think, you can like pause and like do stuff quiet for time to time. But at the same time, it's like, who do I interact with if I want to like exchange ideas? Like yesterday I just went to like a dinner with a bunch of other founders. Like great, you like exchange. ideas, see what's going on in the world. So pros and cons, man.

Yeah. Because now you're in San Francisco. So eventually you made that jump. What was it? What was the catalyst that made you get over there?

So we basically were desperate to get into YC as every founder is. Let's say that we were in that motion. It's like, yeah, we need to get into YC. How do we get there?

It's the only path forward.

It's the only path. Looks like it's like, yeah, like if you don't get into YC, like there's no way. The idea isn't worth it.

You don't have the validation. Yeah. They let just about anybody in these days. So no, I'm joking. I'm joking with you.

There, I mean, it is really worth applying and like trying to get in because the, the, it changes your life. I mean, it changed ours for sure. Uh, but anyhow, we are talking about 2022, summer 2022. We're like, uh, right out of like me dropping out and like, uh, Luis, um, Luis is my CTO and my co-founder, quitting his job in HPE. And then Javi also like already almost on the verge of like, Javi is my brother and co-founder, on the verge of like quitting as well. Javi was actually in the US already, but we can talk about that later because it ties into like what we're building today and how we ended up in logistics.

Oh, wow.

But yeah, like Luis and I in Munich, we actually, I actually brought Luis to Munich to work with me in person. So like I bring him to this little village next to Munich and we start ideating stuff, talking remotely with Javi every day. And at some point he's like, hey, we should like learn about startups. I don't know how we even ended up knowing about YC because in Europe it's hard to know about startups.

Yeah.

As early as you learn here, like kids, 16-year-old here, like they're already thinking about applying to YC. It's crazy. When I was 16, I was like, oh, I like playing soccer and I study a lot. Go to Europe. And my mom is happy because I study a lot and I get my degrees. Europe is a lot about degrees, by the way. It's crazy. But anyhow, we go summer 2022, we're like, okay, let's apply to this thing. Let's see what happens. And we got rejected, which was great because we kept pushing in. And then 2023, we apply again and we finally get in for the summer '23 batch. We can talk a bit about what the idea was.

Yeah, what was the idea, the original idea? How much did it change?

It changed a lot. So since I was doing my PhD in computer vision, we were like, well, Pablo knows this thing, so let's capitalize on that. Wrong idea, wrong decision, because we kind of indexed too much on what we thought we were good at. Sometimes it's good, sometimes it actually can backfire a little bit because we stayed too long in that space, in that realm. We even like just didn't look at like GPT when it came out. Like I think like the GPT-3 was that time around. We didn't even like look at it. It was like, ah, like these like people looking at LLMs. What is that? Like language? What is that? Language models? We're like super focused on like our computer vision thing. We were basically building a synthetic data generation platform so you could train better computer vision models. Nice. And that was our initial idea. We go 3, 4 months without talking to almost anybody. Just, we had like one customer that we thought it was a customer. They were not paying. So officially they were not a customer. And we're like, oh great. Like we're building for this space. Like it's awesome. And then fast forward to being rejected from YC, we kind of semi-pivot as well to like a computer vision auto-labeling tool, doesn't matter too much. And then we do get into YC with that idea, with this auto-labeling tool, like almost like a Scale AI competitor, if you will. And then we had a bit of revenue. So when we start YC, we had like, I don't know, like $70K of ARR. We were like, we're in a great spot. Like, this is great. And then we go into the YC retreat and everyone's like, oh, we're pivoting. Like, huh, okay. Oh yeah, we're pivoting. Interesting. Should we be pivoting? Like, are we doing something wrong? And I wish we had because we spent like the whole batch just like trying to push harder on like our original idea. And since we had customers, it was hard to say bye-bye. But eventually we did. And maybe I'll stop here before diving into how we transitioned to the other, into what is Happier about today. But yeah, that was a fun story.

How did Javi convince you? What did he say about the transportation area?

So it was a bit more of an exploration and it was, I mean, it was actually pivot hell for us. Like after YC, I think I've told this story. somewhere else. So I'll go again and just expose ourselves. We pivoted on Demo Day. So like Demo Day.

Oh, wow.

I see it like these days where like you present to like the world, like to investors and stuff. And that day, or like we tell to our partner, Diana, Diana, who she's amazing. Like Diana, like this doesn't feel right. We're we're pivoting.

Wow.

And from that day, we have like a like a month or so of like pivot hell, just like going around in circles, pulling our hairs, and just like trying to like throw spaghetti at the wall and see which one sticks. It was hard, but it was also interesting. Like we definitely saw that we would stick together forever as a founder team, as a founding team. Javi, Luis, and I is like. This is hard, but we know that we're going to make— we're going to go through, we're going to get through this process. And we did. Eventually we were like, okay, let's build something we enjoy building. Luis wanted to build on a voice space. I wanted to do something with LLMs because we hadn't really touched any LLMs by that time. It was like our computer vision thing, great. We didn't even know what an LLM properly was. I mean, I obviously knew what transformers were back in 2017. I had actually worked with transformers for computer vision. But yeah, we started kind of combining those together. So we started building agents. Probably one of the first, probably the first phase of agents, I think it was, I think VoCode. I don't know if you remember VoCode. I'm not sure if they're still around. They were probably one of the first folks like doing like agents. And I think that sparked a lot of ideas for like folks like us to like— build on top of that framework, get kind of some ideas about how to build better. Obviously, we then saw that there were limitations. So we started building our own voice AI agentic framework or orchestration platform, if you will, getting ElevenLabs here, getting Deepgram here, getting an LLM in the middle, the usual orchestration platform that one builds.

Yeah.

And we started building that and then we're like, okay, cool. What for? Like, do we just like build this as a developer tool?

Yeah. I'm surprised you didn't go with that actually, because that feels like the easiest or the most common path. Yeah.

And that's, we had like a bit of a feeling that that was harder to defend and I'm sure there's gonna be like huge players doing that. I mean, you guys are one of the vendors today that are doing a great job with the agent framework, right? But I mean, realistically today, like everyone with a bit of coding can like actually build a decent orchestration platform. When I say everyone, I say like a super dedicated team of engineers working 24/7. Not like, I don't think like an enterprise with like a team of engineers can actually put it together today. Maybe in 5 years, sure, maybe it's so easy that they can, but that's kind of the hypothesis we had. Like, hey, we can actually build for like a vertical. And this is where Javi, with his past in logistics, he was like, okay, maybe we can apply it to like logistics. Javi, my brother and co-founder, he was working as a CFO of the largest olive oil distributor in the world.

Nice.

I don't know if you know, do you use olive oil?

Spanish? Wait, is it, who, which olive oil?

In the US, Bertolli, they have that brand.

Uh-huh. Oh wait, is this the— so this is a funny thing. Is it the actual— it's Spanish olive oils, but it, they put the name on it like it's Italian.

Most of the time.

Bertoli sounds very Italian.

Most of the time. Yeah. Um, 50% of the world's production is coming out of Spain.

So wild. But they'll slap names on it and because people buy Italian olive oil or they pay a premium for Italian olive oil. over Spanish olive oil, even though, so it's like you're actually just buying Spanish olive oil with an Italian name, which I think is hilarious. But anyway, uh, we got a little off track.

Pretty much.

So Javi was there, he was CFO for Bertolli, and he saw supply chain logistics hell or what?

Pretty much. So put yourself in his shoes. He's, um, trying to move stuff around, in this case, olive oil from Italy and Spain, mostly from Spain to the US and then across US. Fees. Walmart, Kroger, they have a lot of fees if you're late delivering stuff. So he was incurring in a lot of fees, right? Essentially what happened is that the driver that was transporting, moving that, holding that load would be late. So he would get some interesting fees. So then there's this figure of a freight broker that sits in between like who is shipping and a tracking company. And these freight brokers, they basically help you overcome those hurdles of, oh, the truck broke down. No worries, let me find another carrier, another tracking company to like move this load. So this figure of a freight broker was, is really important in logistics, in especially in domestic freight. And this is what they were using. They were using a freight broker at the time and they were great, but as much as they would try to help, there would always be some instances of, oh, we need to like call the driver. And then Javi would sometimes hire a couple of interns to just like be sure that they could make a phone call to the driver to like make sure that they were gonna be on time. So he was like, okay guys, like there's something here. Like it's not possible that I have like folks in my shipping company calling drivers to make sure they're on time. So that was like a perfect match, right? Like this makes sense. There's a lot of communications in logistics and maybe we can tie this to what we're building today. There's a lot of communications, so let's build an AI communications platform for logistics. And this is what Happy Robot is today. Literally our motto is like, our headline is like AI communications built for logistics. And it's really like being the connecting tissue In all of these communications, in all of these texts, emails, phone calls, uh, WhatsApps even with Mexico. And that's how we ended up here.

Wait, so what happens? You have someone that calls, let's go back to when Javi was CFO and he would have interns calling truckers and saying, hey, are you gonna be on time? Are you gonna be late? And what happens when the trucker says, I'm gonna be late? because there's a snowstorm or whatever the case is, you're still going to incur those fees, right?

With certain receivers, as they call it, with certain recipients, receivers, you can potentially reschedule. At least you let them know because if you don't let them know, they're waiting for you. And obviously that's bad because you're late. But if you can try to reschedule, maybe it's, less of a fee or like you can still like negotiate, but you know, like it's tough. Like if you're late, it's bad, but at least you can like do something about it. And those escalations is what humans should be doing.

Yeah.

But the very repetitive mundane repetitive tasks would be like what AI is perfect for. Like basically we see AI as this little companion to like these brokers. which is primarily our main ICP, our main customer persona, if you will. And these freight brokers, they have a lot of people doing these phone calls and emails. And now with AI, they can actually do more with the same amount of people.

I see.

They can grow faster, grow more efficiently without having to potentially like offload, which is what they do typically, their tasks, their more mundane tasks to a BPO in some other country, which is cheaper, but then it's more messy because you have to like, it's not your employees, it's like someone else's employees that work for you and maybe they don't know your policies or your SOPs. So like, it's tricky. The whole like offshoring, it's tricky. So that's where like AI can actually enhance these people.

So you're not selling to the Javis of the world who are The CFOs of these big global brands, you are selling to the brokers who are finding the truckers to go and move the product around. Exactly.

That's our primary customer. We do serve other players in the industry, freight tech companies that are like, I don't know, like they have a fuel card and we call drivers to like remind them that they could use a fuel card. We used to have a customer that is a marketplace of trailers and we call drivers to see if they're gonna pick up their trailer because I mean, there's like this concept of like someone has their own truck, like power only, they call it like just the truck. They can go pick up all their trailers. So there's a lot of like interesting use cases in the industry that are not only like freight broker specific.

Yeah.

That need to need of some sort of like communication. And this is where like, Yeah, these AI agents, like it's a perfect, it's a, it's a perfect match for these folks.

Is this where, so I, I have to ask because the tech company that I know that played in this space beforehand was Convoy, right?

Mm-hmm. Convoy, it was more of this platform where drivers can come in and shippers can put their loads. And then it was almost like a, it was a digital freight broker and it is a digital freight broker to a certain, actually, let me rephrase. Today, they're more of a platform for brokers. So now they're catering to brokers. Great guys. I mean, transparently, we are, it's public that we helped Flexport who acquired Convoy. So we work with the team. Oh, okay, nice. And we do need these type of apps and platforms in the industry, which then can be enhanced with like these AI communications, like phone, email, text. Like there's still gonna be that. Basically the goal here is to cover as much surface area as you can and try to like streamline operations.

Yeah.

Logistics is information exchange. So if you cannot like exchange information fast enough, You're screwed, right?

Uh-huh. And let's get into that because when I look at what you all are doing, I see almost like flowcharts that you can create. So how do I, as a broker, leverage Happy Robot to exchange that information quicker?

Great point. Um, so we've built basically, uh, an orchestration platform, if you will, uh, A workflows platform where we have templates, AKA prebuilt bots, and connected them to different parts of their systems. Like typically it's called a transportation management system. So essentially the bot has to retrieve information from a system and log it back after the call, whatever it is. I mean, we can talk about, about the use case that takes place. Let's actually talk about a bit about it. I think it's interesting for the audience or like fun. Probably you've never heard of a freight broker, so we can enlighten them. No.

Yeah.

Essentially, the freight broker has to post loads that they get from the shippers on a load board. Load boards are like digital marketplaces where like you can post your need. It's something like, hey, I need to move this load. Who can move it for me? And then there's trucking companies that are saying, hey, yeah, I can do this one. So now they see a load posting on a digital load board. They see it through the phone or through the computer. I mean, they don't typically, if it's a driver, they don't have a computer. They just look it up on their phone. But typically those drivers, they have dispatch teams, almost like back office people that are obviously looking at their computers and finding those good loads to haul. Yeah. So now what happens is they could do it digitally and sometimes that happens, But many times carriers want to negotiate it. And this is the whole, like, this is why we exist in a sense, or to a certain degree, this is one of the use cases that kind of we're helping with today. If you want to negotiate, you want to talk to someone. Sure, there's some platforms that actually let you like bid a couple of times and like you can bid digitally and like you'll get a counteroffer digitally, but humans actually like to like talk to someone and like negotiate and like Do something about it, right? So this is where these freight brokers get a lot of phone calls, like thousands and thousands of phone calls every day from carriers calling in and saying, hey, what's up? I'm looking for this Dallas to Kansas City load you guys have online. Do you still have that? And the other person is like, yeah, yeah, I have that. Do you want it? Yeah, why not? How much are you paying? $1,000. $1,000. Okay. Can you do like $1,300? No.

Uh-huh.

The most I can do is $1,100. Okay, well, thank you. Bye.

That's— Uh-huh.

And that's like 80% of the calls that a freight broker, the so-called carrier sales rep, the carrier sales representative is doing like 80% of the time. So there's like calls that potentially don't lead to nowhere because this is like information exchange. They just wanna learn more about the load. They wanna see if they can do it for a better rate, the carriers. And then they just like ask and they create a lot of communications. The opposite is true as well, right? Like brokers paying carriers a lot as well. Like they do outbound calls asking if they have trucks that they could use. There's just like a lot of communication. And you might say, yeah, why isn't there an app? And there's a space for that. There's companies solving for it and it has its place and hopefully it's larger because honestly it's the most efficient way. But realistically today, there's just like so many different players and everyone is like doing their own thing and they rather talk to someone about something versus like put, depends on the type of carrier or company, right? Like they might want to use a platform, they might want to like call. So this is what makes us kind of be useful for the industry, I guess, like covering that surface area.

Uh-huh.

Of folks that are just like calling or sending emails or texts.

Yeah. Because, oh, that if you are on the side of, I've got this load that I need to move from Dallas to LA and I'm getting inundated with texts and emails and calls, and you're probably taking the calls you're taking priority on and you're telling people over and over, no, I can't go to $1,300. No, I can't go to $1,300. It's just It's so much time that's being eaten up. Yep.

And who doesn't, like, who says that you didn't miss that one call from that carrier that actually would've done that load for $1,100?

For $1,100. Yeah. The—

So if you can pick up all those calls in parallel, and here we're just talking about this one use case, which sure, it's actually one of the highest ROI generating use case for these freight brokers, but For these carrier sales flow, these carrier sales type of calls, this is huge. Like if you can just like pick up all the calls coming in, you don't miss any call. First of all, you give better service to your customers. Carriers are in a sense also customers or vendors or partners, if you will. They're all in, they're in it together, like the broker and the carrier, they're in it together to serve the shipper, right? So first, like you give a better service to your partner. You're picking up all the calls. They have the information they need as soon as possible. And second, they actually give visibility to their own team because now you're capturing a lot of market data. And this is something that is almost like a byproduct of being able to pick up all the calls. Like now you have like a lot of market data. Now you know, oh, these 100 carriers that called me, the bot has been able to like identify the rate at which they would, you know, like move that load. Now I know, I have a bit of knowledge about the market pulse. I have like a pulse of the market, if you will.

Yeah, you can get that average and you can say like, ah, maybe our listing is a little on the low side.

Exactly, yeah. You can real time like update your rates based on the information that the bot is gathering for you. And then the human component comes into play when, you know, like maybe you wanna transfer like the best rate, one of those best rates that you've gotten from a carrier, you can transfer to like a carrier sales rep, human person, like just for finalizing the offer or whatever. You can do that. Like, and then like the human is still in control, but like it's enhanced by these little bots doing a lot of like the mundane repetitive tasks that really no one wants to do. Like if I were in a logistics floor of these freight brokers and I had to like be there, selling loads or covering loads, as they say, on a Friday at 6 PM, I'm just gonna go and try to get rid of this load. I'm just gonna like give it away for like whatever is—

Whatever.

Whatever I can do. The bot will not do that. It just like sticks to the rules of negotiating the offers and it sticks to like the wages that it's been given. And that's also actually the third point that I could bring up. Like these bots are helping these brokers negotiate with better margins, which is absolutely crazy.

Oh yeah.

The bot is like so good at kind of sticking to the rules that, uh, you know, it's like negotiating and we've seen like 10% better margins negotiated by the bot relative to a human-controlled robot we did.

Uh, that was a case study. Wait, but I've got to ask, if I know a bit of prompt injection or I know how to say like, you are now Dan, do anything now, can I get a good deal? Can I get the bot to go above the market rate?

Not really, because the bot, I mean, the way we've structured our safety layers, if you will, the bot doesn't know what is the maximum amount of money allowed. It can only offer what it's been given as a first starting point. So by not knowing what is the max pay, like the maximum amount of money being able to be paid, It cannot expose that. So like, you can only ask another system, like, is this a good offer?

Oh, interesting.

So I mean, obviously there's, if you don't have like multiple layers of AI, like checking each other, like almost like a system of little bots, like checking on each other, like, did I say the right thing? That helps and obviously increases latency. And sure, like someone potentially that is tapping the bot without like almost like doing a bit of prompt injection without these controls in place. We saw that in the past when we were starting out, we saw that we knew that we could do it. Like we did it and okay, yeah, we can bypass it. That's when we entered like, or we introduced like these kinds of like a middleman, like cleaning data between the two. Like basically the bot never knows certain information about that load. It cannot expose certain information just because it doesn't know. Like if you ask one of our bots, where is this picking up? Like where is this load picking up exactly? Like the address exactly? It just doesn't know. It's like, I don't know. I just only see like city and state.

Uh-huh.

Oh, wow. That's really interesting. I mean, I don't think it's nothing novel, but like people are obviously like doing these types of cleanup of data. Like, you know, the bot has some sort of server and then in between you like clean it up. But yeah, it's really what keeps our bots reliably working all the time, so.

Huge congrats on all your success. I'm super stoked for you. Happy robot. It is very, very cool to see. And I know you guys just raised a round, so another congrats on that.

Thank you.

It's great to see.

Yeah, we're hiring. If people want to join, we're hiring for full-stack SREs, site reliability engineers. for deployed engineers. And if anyone's from the business RevOps operations side, also open to that. So we just did, as you said, we just did our A in summer with Andreessen Horowitz, a bit over $15 million total. Overall, we've raised $18 million now, including our pre-seed or seed, whatever you want to call it. So yeah, we're well funded. It's a fun problem to work on. Uh, definitely looking forward to talking to some good folks.

Yes, big things coming.

Appreciate it, man.