Building AI for Logistics: A Conversation with Pablo Palafox, Cofounder of Happyrobot

Audio killed the video. Audio is the main thing.

Yeah.

Awesome. So I wanted to start with the founding story of Happy Robot and how you guys— I don't think a lot of people wake up in the morning and say, hey, I'm about to disrupt the world of logistics with AI. So I was excited to know a little bit more about how— what's the founding story behind Happy Robot?

Yeah, it's a— It's a long one, right? Like we, we, I was doing my PhD in Munich and at some point I'm like, okay, let's do this internship at Meta. Sounds cool.

Yeah.

But eventually I'm like, this is not for me.

Yeah.

Let's do something else. And so I, I remember actually the first time I started thinking about startups is I was doing these online masters for like real estate.

You were doing an online master's for real estate? Oh, wow.

Then I go to my buddy Luis, my best buddy from uni.

Yeah.

Like, dude, let's just like buy a house in Spain and like rent it. He's like, No, let's actually build a tool for us to know what house to build, to buy. I'm like, shit, that's good. That's a good idea. Yeah. Like, hey, that's like a startup, right? Like, yeah, that is probably a startup. We gotta make a startup out of that. Funnily enough, the next day I go to the, to the, to the guy that was like building the, preparing the master's. I mean, it was like an online course, right? And I'm like, dude, we have these ideas. He's like, that's crazy. Let's do that. And immediately he was like, organizing things. He was like starting to get us involved in the startup community in Barcelona. This was while I was doing my PhD still in Munich.

Oh, wow.

Yeah, I was, I was still there and Luis was still in Spain.

Yep.

So long story short, we obviously don't follow that, uh, because it was a too small of a market in Spain, right?

Yeah.

But then that kind of sparked the idea to build something cool. So then I, I involved my brother Javi, who had been always like pushing me into doing AI and like computer science and stuff. Javi's my other co-founder. Yeah. Uh, at some point we're like, okay, let's just do something that we know that we have an expertise in. Right. That's when we say, well, I mean, Pablo, supposedly Pablo knows about this thing called computer vision. Right. So why not have Pablo?

So your PhD was in computer vision?

Sorry. Yes. I should have mentioned that. So that was my, that was my PhD. It was in AI and deep learning for 3D computer vision. So like, basically I was like reconstructing 3D shapes out of humans dancing in front of a depth camera.

Oh, that's pretty cool.

Cool stuff. Yeah. Yeah. Really, really cool. During that time, actually, maybe I should tell a bit of that background story about Javi and Luis. Luis is my, again, my buddy from uni. We did robotics together in Spain. We did master's together in Munich. Uh, during that time, Javi, who comes more from a law and business background, finance background, he was actually CFO of a shipper. Now, now you can—

Now everything is starting.

You're going to start seeing the connection soon, right? So yeah, here's, here's, uh, actually in Dallas for 6 years, um, in Diolio North America. If you like olive oil as a good standard, obviously, uh, you might use Vertoli.

Oh yeah, of course.

Vertoli in the US, Spain's like Carbonell and these other brands.

Yes.

So Javi was actually CFO of DeOlder North America. So he was like feeling the pain of actually having to call some drivers sometimes bypassing the freight broker, which is funny. We can talk about that more later.

Wow.

So closing that kind of intro for Luis and Javi, maybe what happened was where we were, right? Like, uh, Let's do something where we have expertise in, right?

Yeah.

Computer vision. Cool. What can we do there? So we started ideating stuff. Eventually we built this computer vision platform, uh, with which we actually get into YC with.

That's awesome.

Uh, so we were like doing some revenue at that point. Like that was last year, uh, March, February, March last year.

Yeah. Yeah. Yeah.

So they, they, we got in, uh, for the summer '23 batch.

That's awesome.

And, uh, that changed everything, right?

Yeah.

It changed our lives. We had to come here.

It changed. Yeah. Yeah. You guys moved, the 3 of you moved here. I think some of you were already in the US. Yeah.

Fabi was already in Dallas. He was working at this company, this shipper, right? And Luis and I were in Europe actually.

Oh, wow.

Yeah. We started kind of ideating in Germany. I actually brought Luis to Germany with me because the ecosystem there was a bit better in Munich.

Yeah.

But truth be told, the ecosystem is not as good as always San Francisco.

Yeah.

Yeah. So, so eventually once we got into IC, like the decision was clear, like we're obviously going to move, try to move here. There's visas and a very long story that I wouldn't bother people with, but, uh, yeah, painful enough to, to, uh, painful, but also like obviously good pain to go through. Like it's obviously going to help you like build your business here. So that's great.

That's super cool. So you guys started more in the computer vision stuff. What was that like aha moment when you started working on logistics?

Yeah. So as I was saying, YC was like pivotal for us. At some point we see that these computer vision ideas are not going to fly. Like we were kind of competing with Scale AI, I guess.

I see.

Or like RoboFlow. Funny, I play soccer sometimes with some folks in here in city, right? And the other day, like I met the co-founder of RoboFlow, which is like a company that I was like, I admired so much. I was like, holy shit, like I'm going to play soccer with you. That's crazy. I love that.

Yeah.

So, so yeah. Close parenthesis. We, we saw that it was not going to fly. Diana and Michael were our partners in YC, Diana Hu and Michael Seibel, and they pushed us to see if we could sell to like enterprises. That's kind of where we saw that it was that build versus buy. So, so we kind of didn't feel right in that space. We tried selling to like, uh, like satellite imagery companies that would use our computer vision stuff to like detect cars and vehicles and whatnot. Government was like there, but selling to government is a pain.

Like, oh yeah, it's super long cycle. And the sales cycle. Yeah. The sales cycle.

It's you.

Yeah.

It was funny. Like we had a demo with an agency and they're like, oh, this is pretty cool. And then it's like, let's rethink in like 6 months.

Oh no. Cool.

Yeah.

Yeah.

That sounds good.

That's a short time in your time span of like a contract lifecycle. Yeah. Wow.

But we eventually, I mean, I don't think I've told in public, but Demo Day NYC. Yeah. It's like the day where you like present your stuff and it's like, okay, well, you're going to get a bunch, a bunch of investors and money.

Yeah.

Uh, actually good startups close before Demo Day, as they say, right?

Yeah. I've seen a shift in the last few years where like companies used to just risk it all on Demo Day. And I think in recent years it's been like a little bit more like, no, no, no, let's do it like Japanese, like boardroom style where like by the time you are there, everything is like, you know, ready and locked and ready to go. And then the Demo Day is more of a ceremony. I've still seen a lot of companies that like by Demo Day, they're like, oh my God, we got to figure this out. And they kill it. And so you never know.

100%. There's, there's, it's not a good proxy. Like the, when you raise is not a good proxy for success.

Of course. I totally agree. Yeah.

Yeah.

Yeah.

It's definitely not. So, but anyhow, we, we, we didn't actually raise a lot, uh, during YC or after YC because we were kind of pivoting. But the funny thing is we pivoted like literally officially on Demo Day. Like Demo Day is like, I presented the whole idea and I'm like, What? So this is our company. And then I'm thinking, shit, I'm just going to like pivot this afternoon, which obviously you're not going to get money for that because you, you don't want to do that. You don't want to get money until you are sure that this is going to make sense. And we already had a bit of money. This is what YC is for. They give you like a half a million to just explore, right? Which is cool. Um, so, so yeah, like that, that was kind of what made us start thinking about other ideas. And now finally, I'm going to go into the logistics connection, I guess.

Yeah.

Yeah.

Yeah.

We started building something that we thought was cool, was conversational AI agents.

Yeah.

That's kind of what Luis and I, Luis is my CTO, uh, very technical, as I mentioned before. Uh, he was like, dude, I want to do voice. So we started building in that arena and then eventually we're like, okay, but for what? Like for what industry? We don't like, like to be too horizontal. At least we don't know how to do that. There's other players that do that great. We don't. Yeah. But Javier was like, okay, well, I know I've been in logistics. I mean, I know. I'm a retailer. I knew the pains in the industry of these logistics thingy here. Like, that sounds like a huge industry to kind of solve, right? And there's so many pain points.

Yeah.

It's like, yeah, there's like tons of phone calls, tons of emails. We're like, perfect. That's a great kind of opportunity for GenAI, if you will. I don't really like that name, but like—

Me neither. For like, we could talk a little bit about that because yeah, yeah, you're totally right.

But let's say like for AI to be applied to, right? So like it's. It's great. It's like also not super high risk. Like it's not healthcare. It's funny. Like I, we actually had these meetings, these office hours with, with Tom Blumfield, one of the partners at YCM, co-founder of 2 unicorns in the fintech world. And we went to his place. He's like, okay, don't fucking touch like healthcare or finance. It's going to be like super painful because of all the regulations. Uh, just make sure that, that whatever you do is like Some playground where you can just break things. And like, he was like, just, I don't know, I moved the other day, maybe do moving companies or like, okay, that also kind of connects with logistics. So like, he was almost like validating our, our go-to-market, our go-to-market, if you will, of like, yeah.

So interesting. And I feel like, for context, when we were building Silo, we, I interacted very adjacently with the world of logistics, right? Because even though we were not building for the logistics world, Um, our users were like distributors and farms and they needed to move stuff. And so Upal, my co-founder, he like built a lot of the like integrations to like, how do we enable these people to like actually move goods without us being a broker or, uh, even a carrier or anything of that sort? And so when you told me, cause I think I met you guys when you were not yet fully into the like logistics part.

Yeah.

But just starting to get there. And I was really excited because after working with some freight brokers, the carriers, the shippers, I came to learn, earn a lot of, like, I had a lot of respect for the people actually doing this job, like every day, because it's a high throughput, high volume work with calls, like, like continuously.

Nonstop.

And I'm assuming you guys are feeling like, feeling like there's like calls to get quotes. There's calls to get status tracking. Everything is conversational. And so I think I'm really excited because I feel like of all the areas, I mean, you were just saying like you didn't want to horizontal, like do horizontal stuff. You wanted to like hone in on an industry. And of all the different industries, I think healthcare and financial services for conversational stuff is a little bit trickier.

Yeah.

I think for logistics, it's tricky, but the amount of value that you guys can provide is incredible. Right. And these guys are like, they have this issue like every day.

100%. When they see that we can help them automate some of the repetitive transactional phone calls that truth be told, they don't really add any business relationship as like logistics is all about relationships.

Right.

It makes a lot of sense, but there's certain parts that It's not about relationships. It's about like, this is a transactional phone call that needs to be taken care of, or this is a transactional email that you need to just address and send a quote for, or just send a tracking link or whatever. So.

That's, that's fascinating. Let's, let's talk a little bit about that, like human. So I think there's been a lot of conversation about, especially in conversational, like, like kind of like interactions. The role of the human being versus the role of, you know, an agent or an agentic workflow. I think GenAI, like you were saying, is a little bit of a generalization that is more about the underlying solution rather than like even the packaged product or the problem, right? The problem is not GenAI. The problem is like, God, I'm receiving 1,000 calls every day. I feel like I read somewhere, I think you guys have discussed before the soul. Of the, uh, of the, of the human in this case, or like the human, at least the human role in this transaction. And you were just saying that logistics is relationship-based. I'm wondering how you feel about that balance between automating, automating some of the most tedious parts of it while keeping that like human relationship sort of active, you know?

I think I can talk about one, one of the biggest use cases in our, in, in, in the In our customer base, uh, which is basically carrier sales, like we, we primarily serve brokers, right?

Yeah.

Uh, we're fully focused on domestic. We have a couple of customers that are like more international, like more on the freight forwarding side. But in freight brokerage per se, there's this concept called carrier sales. Basically, you as a broker, you have a load from a shipper. You need someone to move it.

Yeah.

Need someone to haul it. So you're going to Typically posted on different places.

Yeah.

DAT or DAT is—

I remember that.

The one load board where typically all brokers are gonna post. People are gonna tell you, ah, I never post on DAT, whatever. Like, but people still do that. So like you, you, you would post a load, uh, and then you're gonna start getting a lot of phone calls from carriers.

Oh, interesting.

Just saw that online. They're like many phone calls.

So— Who want to take on that load, essentially.

Yeah. You have a carrier, a trucking company. And I think it's like 95% of the trucking companies in the States is like under like owner operators, basically, like just a guy with less than 10 trucks, like a small company. So there's so many. Um, and that's why, unlike in Europe maybe, which is more, uh, homogeneous, like you have like larger carriers in Europe, I believe, uh, here it's like a guy and a phone and a truck.

Yeah, exactly.

Like literally, and then that's going to create a lot of interactions. A lot of email interactions, a lot of phone calls. He's maybe some carriers have like dispatchers, right? Like, uh, you have like a secretary, like a dispatcher is going to sometimes even help like multiple carriers also.

Yeah.

But anyhow, uh, you'll have like these dispatchers that they saw a load online. They're like, oh, I'm going to try to get this, this, uh, load for my, for my driver. So they call and these brokers, they have like thousands of phone calls every day. That they cannot pick up. And now I'm going to go to your question, right? Like, where does the human connection come in, right? So these phone calls are very transactional. It's like a guy, this dispatcher calling in saying, hey, I saw this load online, Dallas to Kansas City for a reefer. And then the broker is going to be like, sure, what is your MC number? Like the motor carrier verify, like the motor carrier number. Like, who are you? Like, what is your MC number or DOT number? These are like the The, the, the IDs that, uh, that a carrier company has in the US, right?

Yeah.

Like, what, what's your MC number? That's kind of the, the first question a good carrier sales rep should ask to like vet the carrier because there's a lot of fraud in the industry, right? We can also even talk about that. But anyhow, um, it's very transactional. You ask for the MC, you ask for the load number, and that's it. Like now as a carrier sales rep, you're gonna pitch the load. You're gonna be like, okay, yeah, I found the load. This is a Dallas to Kansas City for a A reefer, uh, it's picking up tomorrow at 6 PM, delivering whatever. Uh, I need a, it's a load of shingles. Uh, sorry, if it's a reefer, not shingles, but like, I don't know if it's a flatbed, like you're going to be like, uh, these flatbed needs 8-foot tarps and the driver needs a tweak, like things like that.

Like very— Like the parameters of the load.

Like, what do you need as a driver, as a carrier to pick up this load? I need you to have like your driver maybe needs to have a, like a commercial driver's license or like they need to have 8 load bars or whatever. So for conversational AI, it's very easy to do that kind of stuff, test, because it's like a, like a JSON response from a server.

It is literally just knowledge transfer.

Yes.

Yes. There's nothing relations, relationship focused about that. There can be, there's probably other parts, right? But like the initial knowledge transfer can be, it seems like it can be much made easier for both parties.

Yeah.

And neither party is going to feel like anything is being necessarily like removed from the relationship. Yeah.

Let's continue with the flow of the conversation and then we can also talk about the parts where a relationship could be created with a human, right? With the carrier in this case. But these type of phone call, yeah. Like once we, you pitch the load, you're going to like say, okay, I need like $2,000 for this load. Do you want to do that? The carrier is going to be like, no, help me on the rate, man. Like, can you do like $2,200? Things like that. You're going to negotiate a little bit based on the parameters again, like based on the ranges that the broker can allow themselves to sell these Load 4, you're going to have the AI do that as well. And then eventually what we do today is typically just transfer the call to a human rep, to the actual carrier sales rep, who's also like going to be vetting the carrier again, just to verify, you know, to combat fraud and stuff. And then just like send these rate confirmations, which is basically like a document for the driver to just accept it. So that's pretty much it. Now, where does the human connection happen? Well, It could happen in the sense that, you know, that, that, that carrier, like the carrier sales rep in the broker, of course, yeah. Be like, oh, I know these carriers. I know Pete from ABC Tracking. He's great. He helps us a lot. That's, that guy is already gonna call John at the broker X because they already know each other. Like they, they already have that relationship.

Yeah.

But there's a plethora, there, there's a, a certain percentage of, of the loads. In a given broker, in any broker, there's a percentage of loads that are not going to get covered by your carrier base that, you know, you need to tap into the broader market.

Yeah.

I, if I have a truck, I might be somewhere in Dallas today. Maybe next week I'm in like Washington.

Right. I see what you're saying. So you guys are in, in a way you guys are getting them more business. Like you're not necessarily, I think that's so interesting because I feel like so much of the conversation and this gets me a little riled up. So I'll like summarize it, but And I wanted to talk about this with you. So much of the GenAI conversation has been about the AI coworker replacing the human being. And I think when I see applications like yours, it's very clear that it's less about replacing the human and more about allowing the existing humans to do more business better, faster.

Enhancing them.

Yeah.

Yeah. 100%. We see like 20, 25% more loads covered per person per month. In a broker. That's crazy.

That's the goal, right? It's a tool.

It's a tool for them. They love it. I was actually in, in, uh, in Chicago in one of our customers' logistics storage, logistics floor, if you will. And they're like, this is a happy guy, dude.

I love it.

Like I see the options coming in from the bot. Like it's just like the, the, the AI is going to log the options from these carriers on their TMS, on their transportation management system, right?

Yeah.

Going to log everything. They're like, I love it. Like I don't need to pick up the phone call. I just See what the bot does, and even if let's have the following situation: like you call, you're the driver, you call me, I'm the broker.

Yeah.

You want to pick up at a different time. The AI bot is going to tell you like sorry, this is a strict appointment. You're going to be oh sorry yeah thank you for doing that appreciate the call like I'll call tomorrow for another why should a person pick up that call right?

Yeah especially and I don't think it's like if anything it like filters out the noise and it's even better for the person. Calling in. Yeah. Because you're not wasting their time.

Yeah.

There's no— that's awesome. That's great.

Exactly. And you still log that information, like more broadly speaking, I guess I'm going like too deep into these kind of use cases for logistics, but like more broadly speaking, like you, you still log everything. Actually, you'd log more than you would like, like a carrier sales rep in a brokerage. They never have the time to take a note of, oh, Antonio called us and he wanted to deliver An hour earlier. Never, no one's going to do that. But the AI does that automatically. Like at the end of the phone call, in this case, it's going to extract everything from the call and it's going to log it into the TMS automatically. So now the next, the actual humans monitoring the loads that they need to cover, they're going to be like, cool, actually we can have Antonio deliver an hour earlier. I'm going to call Antonio back and tell him, but they were able to also capture 10 other phone calls that that actually maybe were a better rate. Actually, you know, something crazy, like I think, uh, it's already public. Like one of our customers shared online, our AI negotiates 10% better margins than a human does. Like it negotiates better than humans. And in this case, and again, I'm maybe going to go too deep into this one space, but in the care sales piece, your care sales reps are not fully aligned with the company's incentives all the time. Like if you have a load assigned to you and I have a load assigned to me and I get a phone call for your load because you're busy, I'm just going to try to get rid of that load. Oh, I don't get the commission anyways. It depends on the broker. Like they work differently. Like brokers are different, obviously, but the AI is fully incentivized with whatever company's incentives are. So this is crazy for obviously the, the, the company as a whole. Like the AI is going to be super strict, super con— like it's going to always do the same. It's going to be very consistent. And it's also going to help, you know, those, those human reps to get more out of their day. They're again, they're covering more loads per month, which is good for them because they, they don't get stressed out with calls where it just leads to some— I mean, many of these calls are just saying, sorry, we cannot work with you, Antonio. Your MC number is too young. Like, like you just, like your, your history is like, we don't know who you are yet. We, we need to work with you maybe in the future. Like we cannot work with you today.

So interesting.

That's a phone call that you just wasted someone's time. Yeah. But you have the AI just vetting that.

So in a way you could use Happy Robot also as like a risk management, right? Like to prevent fraud, to prevent to like— I think that, you know, you mentioned something really interesting just now and it's logging. When we were building Silo, we gave a lot of grief to our salespeople because they wouldn't input absolutely every interaction on Salesforce. And now that I'm the de facto salesperson of them, I have a lot more respect and a lot more patience.

Yeah.

Because I just don't have the time to log everything. I feel like BAM and Happy Robot share something in common, which is that we build for mission-critical industries. And so when I think about AI, and I wanted to get your thoughts on this because I think it's AI as an enablement for business processes has so much power. But you mentioned logging and I don't know if you guys have, have you gotten feedback about Hey, this is the level of transparency that I need to your— in your app. How did the AI make decisions? How did it log events? Have you guys thought, because I'm assuming it logs the interactions, but also like, what is the negotiation going on and things like that? Yeah.

I mean, in our platform, you basically can decide what you want to extract from a call.

It was like, oh, very interesting.

Okay. Tell it like what parameters you want to extract. It's fully modular.

Yeah.

And you say, okay, from this call, I want to extract what the offer was from the carrier, what the, what our final offer was, or what the carrier's name is. Okay. This is ABC Truck. Like those parameters you define in the settings of our AI.

That's fair.

And then you can extract it. So like the user defines what they want to log back into their system in a sense.

That makes a lot of sense because not everything matters to everyone at the same time. Yeah. At that level, I mean, that way you almost add another layer of transparency because I feel like So much of, and again, we go back to this whole thing of like AI agents and AI coworkers. And I feel like a lot of companies are building this in a very opaque way where you don't really know what's going on. And if it fails, it fails and you're just like, and so I think the idea that in Happy Robot, you can sort of choose what matters to you and lock that. That, I mean, it's in reality, it builds a little bit of credibility also.

Right. We get that feedback from, from our customers that are like, Okay. This looks like very, like there's a lot of knobs everywhere.

Yeah.

This is like something I would need to learn how to operate properly, but you're giving me already prebuilt. So we have these templates for different use cases in logistics, right? So that's helping them a lot. But what they say is like, at least I know that I could tweak it. The ability for me to like be able to change something is really powerful. Really, really powerful.

Yeah. Give, give me, give them enough knobs and dials so that they can adjust it to their own business. And have you guys found that there is like, is there a balance there between how much complexity you expose to the user versus like, you know, is there like a simplicity versus complexity kind of like, okay.

Yeah. Yeah. There's certain parts that we don't expose today. We just get the feedback from the user and, and we just do it in the backend for certain things. Like, I don't know, we have this concept of What to expose. Actually, this is an interesting topic. Like when you pull up information for a load from a TMS, there's certain information that you don't want to expose to the AI because you don't want to run the risk of it like being, you know, prompt injected or things like that. Like, I don't know.

Brilliant.

So there's these layers of protection that you can build around it, which, which we've done, which is kind of. Guardrailed against someone trying to steal information about a load, yeah, and trying to steal that load. Actually, there's again you know from your days in Zillow probably like there's a lot of fraud in the industry, like people stealing trucks, like literally stealing trucks and like doing all sorts of crazy stuff, like like double brokering is a big one as well. The reality is that you don't want to expose certain information online. Yeah, actually that's why many people don't want to post on these load boards.

Yeah, I hadn't thought about it actually. Oh wow!

Like the shippers tell the brokers like don't post my load on on these load board because I don't want it to be like public.

Yeah.

Obviously you need to publicize, publicize it somewhere so that people see it. Otherwise you're not gonna get it covered.

It's so interesting cuz there is enough parameters and like if you load, if you go to a load board, there's enough parameters in there to like do some harm.

Yeah, there could be, especially if you like, obviously if you know the shipper's name, you just, you can just go there and obviously there's things around it. Like, I don't know, like, like a broker will never give the pickup number to the carrier until they're really in the shipper. So there's like certain verifications that you can do, but brokers are like, sorry, like bad people are gonna, you know, get creative and we're just getting, we're just getting some noise.

This is an amateur operation here. Like we're getting all the probably cut from the street.

Probably this is gonna add to the realism of like, will people think we're like AIs or something?

Yeah.

Um, no, like what I was, where I was going is you expose what you want to the AI. And this is something that we can also talk about, like how Or what an AI agent should have information or have access to.

That makes so much sense. So what you're saying is you're not shoving everything into the context window because that's not safe to begin with. Yeah.

Like, why would you, right?

Yeah. Especially if people are going to be interacting with these systems from an external capacity, you don't want them to be able to have like what would be considered a modern type of SQL injection, right? Where like, You're probing the AI.

That's so interesting. So there's, I mean, there's different approaches. I'm pretty sure like eventually in like 6 months from now, the state of AI agents or AI language models or LLMs will just be different. So maybe we change our approach, but today it's that framework where we can just decide what to show to the AI.

That's brilliant. And I think that you can do. I've seen a couple of strategies being used in the, in the last few months. One of them is obviously a lot more deterministic, which is like masking, um, and, and things like that. And the other one is a little bit more into the actual, like, how do we build models that automatically remove certain, like, delicate information from the underlying model? I know Wombo Labs, for example, has been working on, like, a way to basically eliminate automatically things that are delicate from the training datasets so that you don't leak Things, especially if you do fine tuning. What are the challenges? I mean, and I think that the reason why I was mentioning before, like, I feel like something that, like when you work in building in a mission critical industry, I mean, setting aside AI, just building software, period, in a mission critical industry is hard. And it's also the kind of thing you guys have built the kind of product that if it stops working.

Yeah.

Then you start getting phone calls.

We do.

And that's, and that was the pain in the beginning. Yeah. And that was the pain of silos as well. In the beginning, my co-founder and I would just receive phone calls at 2 AM. Also, these guys are working through the night. And so we, uh, that prompted us to like work on reliability very, very early.

Same here.

Yeah. Because we all wanted to sleep through the night, which was, which is great. Sleeping is great. And so I'm wondering if you felt, I think As we build, because we are building AI components that are much more horizontal, but I will say though, most of our customers are in like pretty mission critical applications. That's a conscious decision we made from the beginning because we wanted a hard, hard problem and add more value. But I also feel like there are challenges involved. So I was wondering what challenges have you guys seen? This is not just some. AI chatbot that you can put in like an e-commerce website and all the respect for AI e-commerce, but like, it's, this is critical, right? You can lose money if you don't. And so I'm, what challenges, especially when working with these like statistical abstractions, because at the end, you know, and you've done this in computer vision, like you're calling a function that may be a function, but it is, it's a statistical function.

Yeah, there's chances for error.

Yeah.

Where could I go in these, with this question? So basically the, let's start with what we've built around the infrastructure, I guess. One of the biggest, biggest pieces we've done is making sure that there's always a fallback.

Got it. That makes sense.

So, I mean, there's nothing new. Everyone building with kind of any sort of LLM is going to be like, okay. If these API requests or if my GPU is down, I need a fallback. So thankfully today there's so many alternatives. Like if you're self-hosting your LLM, like Llama 3.1 or whatever, you can also maybe have like a fallback server with like another instance of that. Or you can just directly fall back to like maybe another LLM that is—

Like a Together or like a—

Exactly.

Yeah.

Now, if you have like, I don't know, like a Claude or like a GPT-4o, whatever, you can I don't know, like do one, like the Azure RAG and then like also rely on OpenAI. So like these fallbacks we've built across our different models. We, we, we do support different models, right? Depending on the use case.

Yeah.

So that's, that's a very important one. We, which we now have figured out pretty nicely. We still have to like improve it, but obviously it helps a lot because now the system should never go down.

Right.

As opposed to like the first days of Happy Robot where we were like, Shit. Like they were actually waking up in the morning.

Like things do go down, right?

That's Pablo. Like they're actually calling all 3 founders, like, guys, like this is down. Uh, and obviously we're like, we built around it and now it's like, well, now we can sleep. Yeah. That's good. Especially because there's a lot of moving pieces in, in like, uh, like especially voice AI, there's a lot of moving pieces that you need to make sure everything works. Not only the voice per se, like the voice AI per se, but also like the connection to their TMS. Is it logging? the information, right? Is it sent like, sometimes we send like an email to the rep on a load or whatever. Like we send an email notification or we send a text during the call or we send it like the agent can do things during the call, right? It can only speak also like retrieving information, but also like send a text or send an email. Like those processes are like, you know, they, they, they can all break. So it's a lot of moving pieces. But I would say having those fallbacks helps a lot. Yeah, I would say, I would say that's where I would, I would go here.

No. And I think that it makes total sense. I mean, I feel like you guys have built the infrastructure for resiliency, right? Especially given that these are stochastic processes. I also feel like you guys have an additional layer of difficulty on top where you are literally, this is real time, right? Like that's the thing. And so this is not a deterministic, like this is not an asynchronous process. As much as it is something completely has to be live. So latency matters. The user is going to notice it.

Latency matters a lot. If you, if a request, like if your transcriber, like we have in the, in the phone call, you have like something transcribing the call, right?

Yeah.

If your transcriber fails for whatever reason, you need to have like a way to know and like very quickly, like have a, like a support transcriber service. doing that because when you're like processing like thousands of calls every day, it gets crazy.

Like, no, no, of course you need to learn how to scale that.

And then yeah, like the latency piece is huge. Like if you don't respond quickly enough, um, it's going to create some weird situation there with the user, right? Typically the user doesn't even know it's speaking to an AI most of the time, which we—

That's incredible.

Yeah. And we can actually talk about that. From a, from a, I don't know, like not a legal perspective, like from a moral perspective, like should, should one be telling the other end that it's an AI? I think eventually we'll all be talking to AIs all the time. So it's going to be like buying on Amazon, like 10 years, 10 years ago, like my mom would be afraid of like going to Amazon and buying something. Now it's like you just put your credit card anywhere. So I think we can also, yeah.

I also think though that, 'cause I've heard of this, I feel like there's been a lot of conversation about The disclosures on AI. And I feel like for your particular use case, like you guys are not doing like robocalls to like, like for campaign, like, like detection or anything. And so this is very much a business use case. This is awesome. What are, what is next for Happy Robot?

Growing, growing a lot, serving our customers. I mean, I, it's crazy. Like before starting with some of the customers we have today, I was like willing to just go to like a logistics floor and like learn more about their pain. Now I actually get to solve it today. And it's really cool. Like I really like the industry. Like it's, it's really a cool industry. Like it's really fun. There's a lot of things going on there. I'm going to a conference tomorrow in Charlotte. Uh, next, in 2 weeks from now, I'm going to another conference. There's so many things happening and people really want to like improve things in the industry. Um, there's definitely some laggards out there that just don't want to change things, but in general, like there's a drive now to like automate many processes that can help everyone make one of the arguably the most important industry in the, in the world, like better.

So that's crazy.

That's, it's so important. Like if, if you don't deliver a truck, maybe you don't get your microphone.

No, no, it's true. And I feel, you know, what's really funny is that before starting Silo, I had never asked myself when I go to Whole Foods and there is like an apple on the shelf. How did this get here? I think most people don't think about it too much. They're like, yeah, I know there's distributors and I know there's a farm somewhere, but they like, they're not aware of the sheer operation that goes behind it. And it's all sort of purposefully transparent. It's purposefully simplified for everyone. 'Cause like the customer shouldn't have to worry about it.

Yeah.

Um, but I'm really excited that you guys are helping the world of freight because I feel like Having worked with a lot of these freight companies before, I feel like you guys also entered the market just at the perfect time because I feel like a few years ago people were still like iffy about technology. And now that freight, because you remember freight had a boom back in like— COVID, during COVID And now it's not such a boom anymore. And that's actually good for people that are in the industry of enabling people. And I think that. That's a great opportunity for you guys.

It's a down market for, for freight right now. People are saying, are calling it like one of the worst markets for freight these, these days. The coolest part is that we're one of the only pieces of tech that all these brokers in these days are willing to pay for. We have one customer that, that was telling us like, dude, I see like so many pieces of tech like yours, the only one that actually like Helps because there's a lot of noise and not a lot of margin in this industry. So they have to be very, very selective on what they invest in. Wow. And we're actually in a very good spot because yeah, like these, a couple every week, we're like onboarding a couple of free customers or 4. So like, that's growing really quickly. And, and all of these folks are like, holy shit, like this actually rocks my world. Like when we show them that they can automate many pieces of their, of their processes, like in a, in a minute.

Yeah.

I was like, click here, connect here. This is now working. I just can do my check calls now. I was like, yeah.

That's the dream.

And now they're actually like the smart guys are like already preparing for the market that will recover because, you know, like there'll be, there'll be more, more, more phone calls to handle, more emails to respond.

Yeah.

So like, they're almost like preparing for what's coming. So that's really cool. The fact that they are already willing to pay today for like such a technology in preparation also for the next. Upcoming market is crazy. It's really cool.

What are some, some of your personal insights so far? You know, while building CopyRobot, all of this transition, you obviously moved from Europe to here, which is incredible. What's your, what are the biggest takeaways for you?

I like to focus a lot in one type of problem.

Yeah.

I think that helps me be at ease at least. And my, my co-founders are kind of aligned in that. Um, I think that's one piece of advice that I would still give myself again. I think the second one would be like, like the first one would be like focusing a lot in that particular problem. And YC says that a lot, right? Like a startup can only do one thing at a time well enough.

Completely.

The second piece would be, be open to pivoting actually.

Oh, for sure. I mean, you guys—

Be very reactive to the market.

Yeah. It's like these like 5-degree angles is what what allowed you guys to be able to get where you are now, which is incredible.

100%. Yeah. Combined with that resiliency of like, don't give up, have a team. You definitely, definitely would have not been able to make any of these without someone on, on, on our side. Like, same for my co-founders, right? Like, we had each other. We had each other and like, we went through dark moments, right? Like, pivoting is horrible. Like, when, when we were like going away from our, from our computer vision thingy, we had customers to fire. Like, we actually fired customers. It was like, What the hell are we doing? Like we went to YC and now we're like pivoting. But, but it helps come at it, come at it with a, with a fresh mind. Like be like, no worries. Just think about it more broadly. Get, get excited about other things and you'll find something that you're excited about. So I would say like, have that open mind as well. So like that resiliency and open mind with the ability to pivot quickly, be very— I think that's the key. And I think I'm just like honestly reading out in my mind like that, that the 10 or 20. Uh, bullet points that YC had in like a piece of paper. I actually printed out in my, in my place. So that's pretty much it. Like, just, just go for it. And especially the go for it. Like there's many people that, especially in Europe, right? I mean, we come from Spain. Uh, Europe is like so conservative.

With respect, especially with business and startups and technology and—

Yeah, pretty much anything, right? Like everything moves slowly, right? But, but here it's like, just, just go for it. Just like. I don't know, quit your job, start thinking about stuff and build.

Scary. It's very scary, but I feel like that's why you guys are where you are. It's because you had the guts to pivot, fire customers who no one wants to do that, but sometimes you have to. I just realized the other day, I, that you and I are both Politécnicos actually. Yeah.

Yeah.

From, I think we talked about it and then I forgot. That's so awesome. Well, thank you so much, Pablo. This was great. That's happyrobot.ai.

HappyRobot.ai, LinkedIn. I'm relatively active there. You can, people can, can click on some posts and, and see what our AI sounds like. Yeah.

Awesome. Well, thank you so much, Pablo. This was great.

Appreciate it.

Thank you for your time. Thank you.