Automating Annoying Phone Calls in Freight with Happyrobot's Pablo Palafox

it's, um, something that you don't maybe fully know about, it scares you, right? I have to be clear, like, AI is just statistics. It's no magic. It's just, uh, a bunch of statistical predictions. Like, it's just predicting what is the most probable next word, uh, that I should be like saying next. So like, AI is kind of dumb if you think about it. Obviously, I'm just like making it feel like very stupid, but I mean, it's super smart. But it's smart because we— it's in a lot of, um, a lot of data. But under the hood, it's just statistics, right? So nothing to worry about unless we use it the wrong way.

Welcome into another episode of Everything Is Logistics, a podcast for the thinkers in freight. We are proudly presented by SPI Logistics, and I'm your host, Blythe Brumley. We have another great topic centered around AI and logistics tech for you today. We are talking to Pablo. He is the CEO over at Happy Robot, and this is a conversation I've been wanting to have for a while. So Pablo, welcome in finally to the show.

Thank you so much, Blythe. I'm super excited to talk about everything logistics and AI.

Heck yeah. So before we dive into, you know, what you're doing at Happy Robot, I'm curious as to what your background is. Did you know about logistics? Did you work in it beforehand before you started working in it now? Or how did your journey end up in the logistics industry?

For sure. My background is actually in industrial engineering. So I did touch upon some logistics during my bachelor's in Spain. I studied this undergrad with my buddy and CTO now, Luis, one of my co-founders. And basically I then focused on computer vision and computer science in a sense, like AI. I did my PhD in Munich. Before that, I was doing my master's as well in Munich together with Luis as well. So it's been basically a lot about like AI and computer science when it wasn't even a concept. Like I was actually doing some AI when it was more like deep learning. We referred to it as like Like, people who actually said AI was like, this is not AI. And right now, like, people are actually saying, oh, this is AI now, which is kind of debatable. But the correlation with logistics is more through Javi, my brother and co-founder. He was on the shipper side during his time as CFO of an olive oil company. He was the CFO of North America. They were shipping olive oil all across the US and North America, right? And he was basically, you know, dealing with the whole Hey, let's call the driver. Oh, they're waiting on— they're in detention. They're just— something went wrong with the warehouse. Like, that kind of came through Javi.

Oh, that's super interesting, an olive oil company, because that's all the rage on, you know, some of my social media feeds right now is what kind of olive oil to buy that is actually like the real olive oil and not, you know, some, you know, industrial Americanized version of olive oil, quote unquote olive oil. So that's super interesting. I'm curious, though, for your Your PhD, you said you got it in Munich. And so when you're studying there and you're studying computer science and AI, you know, deep learning, machine learning, I imagine all of those things. I'm curious what your thesis was on.

Well, first of all, don't get Javi started on olive oil. He'll like tell you everything about it. So don't even ask him about like what bottle you should buy. I guess I'm glad he's not here, although I would love for him to like be here, like to actually now drill you down.

Now we're gonna have a follow-up.

Yes, let's do that. Just olive oil. But, um, but yet my, my thesis— I actually dropped out of the PhD officially. I never fully graduated. Um, I can go down that path and like tell you about how messed up the like research world is in terms of like just a battle for papers and just like, oh yeah, getting more papers and research. I mean, it's just, uh, I love my professors. They were great. I really appreciate everything I learned from them, but the whole research arena is just meant for you to produce papers. And that wasn't really for, I guess, entrepreneurial mindsets like, I guess, mine. I really wanted something else. I did an internship at Facebook, Facebook at the time, Meta now, I guess. And I realized that, hey, this is not for me, this whole research arena, even if it's through academia or through industry, sorry, through some Meta, Google, whatever. still like just producing papers. So that wasn't for me. And then again, my thesis, to your point there, it was more related to like 3D avatars. So like deep learning applied to like just reconstructing people dancing in front of a camera, just like kind of for that virtual clothing, you know, like Meta is now actually diving deeper into that as well. So that was my thesis, yeah.

Yeah, it almost sounds like—

It just naturally, yeah.

Oh, I was gonna say, it almost sounds like the Web3 that when it was at its peak and I think Meta tried to jump in that arena too, in making the meta-themed bodies for meetings, I guess.

Exactly. Yeah, it'll come. It'll come. Like, um, being able to like just see yourself with different clothing and just try things on, like that virtual try-on, I think it's very powerful, but it's very hard.

What makes it hard? Is it, I imagine, all the different body styles, you know, shapes and sizes, things like that?

Reconstructing your body in real time. being able to like actually and accurately capture your like movements and like your positions and you're like, uh, positioning clothing, virtual clothing on your body, um, in a realistic way is hard. And obviously if you're like trying something, you want to feel good, uh, so it shouldn't like look messed up. Like the image that you see of yourself, like if it's funny looking, it's like, I don't like that. So obviously you're not going to deploy anything that's Making you look weird, right?

Yeah. And then the fashion brands probably don't want to invest in it because their clothes maybe don't look the best. But I'm sure, you know, as you said, it's, it's, it's coming. It's just not exactly, you know, quite there yet. So, so how did you go from, you know, sort of discovering that you're not— you don't really like the atmosphere of the, you know, generating research papers, which is probably the way they get funding. How do you decide, okay, I'm not— that's not my path anymore? I'm going to go this other path? Did you start up a company over in Europe or did you move to the States and then choose to start up a company here? What did that process look like?

Great point. Javi, my brother and co-founder, he had already been living in the US for a while and he's always been trying to distill into me like, you should do computer science, you should do startups. He was kind of waiting for me to graduate and these kinds of stuff, do something together. And at some point, once I'm done with the PhD, stuff. I bring my buddy Luis as well and Javi in and we're like, hey, let's do something together. And we started ideating back when I was still in Munich. We started ideating and building some stuff around computer vision at the time. At some point we get into Y Combinator in 2023, like last summer, 2023. Y Combinator is this very prestigious accelerator based in San Francisco who's funded companies like Airbnb, Dropbox, Flexport. And we got into YC. We actually started ideating about like the concept of like LLMs and conversational AI. And this is kind of how we got into building these voicebots for logistics.

Mm-hmm.

Kind of the connection with Javi saying, hey guys, I've seen a lot of issues with like very transactional phone calls in logistics during my time as a CFO of a CPG company. And we were like, hey, we like AI and voice AI stuff. Let's do something in the middle, right? So that's kind of how it all came together.

So how, I guess, because YC is just something that if you follow any kind of All In podcast or any of the, I guess, that circle of social media influencers, you hear about YC a lot. So I'm curious as to what did that process look like to decide, okay, I want to start a company, but we want to go through YC?

We, for some reason, we always knew that we were going to apply to YC. And actually, to your point before, like, how, like, did we build a company right off the bat? We actually didn't, but at some point we did incorporate. And the place where we actually incorporated was the US, even though we were kind of partly in Germany, partly in the US, right? Like, Javi was here, Luis and I were in Germany. The reason we did that is because it's so simple. This is a company in the US versus Europe. There's so many hurdles and issues and regulations in Europe, Spain, Germany, pretty much all the same. Probably Spain is the worst place to like build a company anyways. But even in Germany, like it's a lot of procedures. Language is also a barrier, of course, like you're having to sign documents in German. So it was clear from the get-go that we were just gonna try to get to the US. Obviously visas were an issue at the time. Now we finally got our visas and we're like being able to like be here and build a company and just focus on that, right? So really appreciative of of the US in terms of how easy it is to build companies. And to your point before, like, that's kind of what we did. We just incorporated as a C-Corp Delaware. 2 days, we have a company. And then when we got into YC, which they actually, actually kind of force you, like, they don't accept you if you don't have a C-Corp Delaware. So like, you actually have to have that. So once we thought that we would apply to YC and we get in and we're like, Great, we just might— we might as well just move to the US, right? So we did that in summer, like last June.

Oh, wow.

And yeah, it was fast. Like, it was kind of like in May we got accepted, June we're like here. And we just had to like bring our lives here. I actually asked my wife to like just find a job in San Francisco or try to get some. And she actually found an amazing job as a school teacher, as a learning specialist here in San Francisco. It's like so easy to just like come here and like do stuff and build cool products. So yeah, that was kind of the journey. We can talk more about it, but yeah, absolutely.

I'm just curious as to, you know, is it just a fund? Are they— I say just, um, but are they a funding partner and also kind of like an educational camp, you know, collaboration group? Like, what does that look like once you become accepted?

Yeah, they fund you with half a million dollars. That's kind of the ticket they give now to startups. But that's kind of just one small part of the whole package, if you will. The biggest and more interesting kind of learnings and kind of benefits you get from YC is definitely the network.

Hmm.

Being able to just like look up online, like, okay, who's the CEO of Flexport and what's his email. Oh, okay, Ryan. Okay, let me get his email. Let me just email him. And he might actually respond because, you know, or he knows that, hey, this is coming from another YC company.

Hmm.

So there's that level of trust and that level of, you know, trusting each other, being good to each other in that community that makes it very special, right? The fact that you can just move across the network, benefit from others, getting help from others, helping others as well yourself. Like, That's actually something that I try to do. And Luis and Javi, as co-founders of a YC company, we try to do ourselves. Like, how can we help other people just applying to YC? I just had a couple of friends from the past from Germany, like they're gonna apply tomorrow to Y Combinator's next Summer Batch.

Oh, cool.

And they're like, hey, can you like give us some tips here? Can you refer us? And of course, like that's something that you can do. Like if you know people that are good and building cool stuff in the startup space, in startup land, and you wanna have them in your network, That's kind of one thing that you can give back. You can just refer them back to that organization in a sense.

That's super cool. What other tips would you give a business that's interested in joining like a YC?

I would say being extremely focused on the customer and on building quick. So the motto of YC is build something people want. So that already reflects and talks to the fact that, hey, you have to be super mindful of the customer. Like, who are you building for? Why is it useful? Why are you doing this? Like, it's just not building a cool product, which is, granted, something that we all do. Like, we always, especially engineers, we always like to like first build and then talk to users. It should be the other way around.

Right.

That's something that we've been trying very hard to like avoid, especially in these, in the logistics space, right? Like, when we came into this space, I didn't know that much about it. Javi knew, but Luis and I, we didn't. And it's been just like being very open-minded, very humble in terms of, hey, you don't understand anything about the industry, you better learn it and you better get close to people that teach you about it. So that's something we've been trying to do. And that kind of talks about how YC or what YC is looking for, like people that are actually listening to their customer, listening to their users, talking to them constantly. Really being, really serving your customer. That's kind of the key thing that you should keep in mind if you want to apply to YC.

Are you in freight sales with a book of business looking for a new home? Or perhaps you're a freight agent in need of a better partnership. These are the kinds of conversations we're exploring in our podcast interview series called The Freight Agent Trenches, sponsored by SPI Logistics. Now, I can tell you all day that SPI is one of the most successful logistics firms in North America who helps their agents with back office operations such as admin, finance, IT, and sales. But I would much rather you hear it directly from SPI's freight agents themselves. And what better way to do that than by listening to the experienced freight agents tell their stories behind the how and the why they joined SPI? Hit the freight agent link in our show notes to listen to these conversations Or if you're ready to make the jump, visit SPI3PL.com.

And I imagine that, you know, with Flexport already being, you know, somewhat of their funding coming from YC, I imagine that, you know, that group is already familiar with logistics. So there's no, you know, sort of convincing them of the value of the investment, which we've seen other, you know, VCs follow suit since. over the last couple of years. So it's interesting that they were maybe early to the game in that regard. Is that a safe assumption?

I would say so. They don't have a thesis. Like, YC doesn't have a thesis in terms of industries. They're pretty much industry agnostic. They would just look at whether what you're building is useful for someone. And if it's a big enough market, obviously, but logistics is huge, right? Like, there's so many problems, so many inefficiencies that we can tackle. So yeah, 100%, they're not biased towards any particular trend or whatever. Obviously now AI, it's not real. I would not even say it's a trend. I mean, it's something happening. It's a technology the same way that mobile had its time, right? So they are very focused on AI right now. They are kind of getting a lot of AI-focused companies in their new batches, but it's simply because I was listening to one podcast from YC the other day. It's simply because they just happen to fund people, smart people that are building stuff with AI. And that's kind of their thesis. Like, their thesis is, I will fund people that are smart, have energy, and want to build cool stuff. And depending on the time of, like, on the decade, I guess, they will be using a different technology, right? Like, now it's AI.

And so you go through this whole process, you move your family over from Spain or Munich, as you said, over to San Francisco and you get accepted into YC, what do those first few steps look like? Where are you, I guess, investing that money that you get?

YC's founder and not president anymore, but YC's kind of creator, Paul Graham, told us in one chat at the end of YC's Our YC's badge. Sit on the fucking money. Just like be a cockroach.

Love that.

And this is something that he's already, he's also like posted about on many of his posts in the kind of startup community programs, like these big figures. And he's very conscious of how money's just a tool to get you somewhere. But if it's not a blocker to just hire someone, don't hire them. Like don't just spend the money because you have it, right? Just sit on it. until you know that there's a path to like become, uh, to do something useful. Like keep understanding your users. Again, going back to like talk to your users, understand what they need. So that's actually what we did. Like we just kind of sit on the money. We've spent like 5% of the money that we got from YC and other investors when we raised our pre-seed. So we mostly have everything in our bank account. Obviously you're now getting revenue from customers, so that's also gonna— like that's something VCs never actually tell you. Unlike YC, like, VCs always will want you to like get more money, like get more money, get more money. It's like, obviously they want that because they're just gonna get more ownership, right? Which is, hey, it's a something, it's fair, like it's their job, right? But what you want as a company is just to be mindful of your runway, obviously, and you wanna have enough money to like just survive and until profitability, but also know that you're gonna get revenue. And this is something that YC also tells us a lot about, like, hey, It's not that you're just gonna have these amount of money, like, I don't know, $2 million or just that's it. Like you're not gonna get anything in. Of course you are. Like if you do something interesting for some people and users, you're gonna get revenue, right? So that's gonna offset part of the costs, right? So yeah, what we did to your question before, what we did is just continue building, talking to people. We actually did a pivot after YC. We didn't apply to YC with this right exact idea. It was a bit different, which is actually something that we can talk about as well. Like people applying to YC, I would say last batch or our batch, 50% of the companies just pivoted during YC, which is crazy. Like they just applied with one idea and they just pivoted. So what we did is just like sit on the money, keep talking to users. And right now we're on the path to actually building something very cool. We are now seeing that someone actually wants what we're building, unlike what we were doing before. And now we might just spend a bit more on like new hires, Uh, more compute and this kind of, this kind of thing.

So it's really like the approach is, is that it's almost more calculated now versus kind of what the VC environment was for the most of the last 10 years, where it's just like fund everything and who cares. I mean, obviously they— people, people care if they make money on it or not, but the money was sort of free-flowing. This sounds like, you know, very like conscious shift where we're gonna give you the money, but we want you to spend it wisely, sit on it, as you said, and then make— talk to customers and then choose to make your investments after you reiterate on the product.

100%. 100%. Yeah. There's that. I think there's been that mindset towards profitability and towards revenue. Which I think is very valid because how do you validate if someone wants what you are building? Well, they pay you. If they're not paying, like, what's the validation?

Right.

It's like, oh yeah, this is very cool, but I'm not gonna pay. No, like, you need to get validation through, like, people paying for your stuff, right? So that's kind of something that actually YC has always been very, very persistent about, I guess. in all their videos, like if you go 10 years ago, 10 years back, you'll see that in their videos, they'll like just be talking about, hey, build something people want, build them, like make them pay for it. Like they're using it, so they better pay, right? So yeah, I think there's that mind shift towards building things that you can actually get paid for.

And so speaking of, you know, getting paid for a product, you know, obviously you're CEO of Happy Robot. You saw this need in the market and along with your co-founders as well. And so how did— what did that process look like? You know, are you talking to customers every day or, you know, are you starting off with a couple of customers and then kind of proving it out and building it out from there? What did those sort of early, like, I guess, first 6 months look like?

Even before going there, I guess I'll give a primary in what we do. So basically, Haprobot, we build voice AI assistants. So like voicebots for logistics. So we're trying basically to automate all of those transactional phone calls that no one is building relationships through. So anything like a payment status update, be it like a factoring company calling in a broker, hey, any advancements, any deductions on these, on this load, whatever. Like those things are very transactional. Who's built— I always hear about, oh, like relationships, relationships and logistics, that's very important. Who's building relationships there? Like no one, right? So like all those transactional phone calls, we want to help companies be more efficient on those. So we're serving freight brokers. We're working already with some of the top 50 US freight brokers, including some like Circle Logistics. We're also serving carriers, automate their dispatch teams, giving load updates to brokers as well. So we have this cool platform that can just basically allows you to just like build visually how your assistant will operate, kind of like a phone system, like a phone tree, but with AI.

Hmm.

So this is kind of what HappyRobot does, like building these voice AI assistants. Now, to your point about how often we talk to users, well, we talk to them every day. We have Slack channels with them. If they don't have Slack, we just— I just give them my phone number and we just talk through texting or like email. So I try to like be super on top of our customers.

Nice.

And I think that's kind of what I wanna distill into my company moving forward, like customer-centric, right? Like just be mindful of, hey, we have someone in production, it better work. And we need to make sure that it's working all the time. 99.9% of the time it should be working and the assistant is doing its job and the customer is loving the product, right? Which so far we've been lucky enough to not have any churn or people just leaving us.

Heck yeah. I mean, knock on wood for, for that statement right there, because that's definitely something to, to be proud of, is not having any of that churn so far. And I think you actually have a demo that you're able to show us, and I think that that will help the audience sort of understand, because it— you— I've been shown the demo before at a conference, and I was pretty much blown away that that wasn't a real person that you're speaking to, because it sounds very much like you're calling into a real, you know, sort of brokerage floor. You can hear the action going on in the background, and you can have have what feels like a good real conversation, but it happens to be a bot. And so I think you have a demo that you can show us, right?

Yes, 100%. Let me just call here. So I'm gonna be demonstrating how our kind of inbound carrier sales rep sounds like. It's a pretty simple flow. It's just imagine you're a— I'm gonna be kind of the carrier asking for loads. And we're just going to be getting our assistant, in this case Kate. She's going to give me some load information about some loads. Let's see how that goes.

Circle, how can I help?

Hey, how's it going? I was looking at your Dallas to Kansas City.

Sure, can I get your MC number?

Yeah, it's 843818.

Let me check that MC real quick.

Okay.

Let me transfer you.

Okay, thank you. So here what we saw is the MC validation was right. and we just immediately want to transfer to someone, uh, to the actual carrier sales rep so that they, they can check on that for you. In this case, the AI saw that the load was not there, so we were like, okay, let me, let me transfer you like immediately so that you can talk to a carrier sales rep. It's like this validation of the MC is already very powerful because like you're only, um, letting people work with the carriers they want to work with, which is already like a huge ROI for many, for many brokers out there.

So you're cutting down on those phone calls of just, you know, not necessarily like spammers or scammers or anything, but just unneeded phone calls for what you're currently— what your workload demands for that day. So she was actually transferring you to a real person or to another, you know, sort of, I guess, robot?

In this case, to a real person, because that was kind of the flow that we intended. But the the potential to transfer to another bot could be there as well. Like if someone wants to like get a load update, actually, uh, let's say for a dispatch team, uh, that load update we will be transferring to another AI that can like give that load update. Kind of having the router AI assistant first kind of routing to the right department, and then within each department you can think about having like different AIs doing different jobs, like one for accounting One for load updates, one for capacity requests, kind of that, that flow. And for brokers, the same, right?

Brokering success demands a battle-ready strategy. Tye TMS equips freight brokers with the ultimate battle station for conquering a tough market. With Tye, brokers gain access to a comprehensive platform where rate intelligence and quote history converge on a single screen. It's not just a page, it's a strategic command center designed to help brokers win. TAI equips your team with all of the data they need to negotiate with confidence and allows them to communicate directly with carriers and customers from a simple control base. Revolutionize the way your brokers perform by giving them a competitive advantage with TAI TMS. For more info, go to tai-software.com/battlestations, and we also have a link for you in the show notes to sign up for a demo.

Yeah. So I just pulled up your website because it kind of breaks this down really easily where it talks about those different calls and those different workflows that you can work through. And I, what I'm wondering is, I guess, are these like maybe like prebuilt filters or prebuilt workflows that you kind of have already established based on those conversations with your customers? And then from there you can kind of customize it. Is that a safe assumption?

Exactly. We have predefined flows already for brokers there. For instance, like those outbound check calls. We actually have some LinkedIn posts coming up soon where we'll like demonstrate how to like do outbound calls, like drivers for in-transit check calls.

Oh, interesting.

Also even for like pre-trip check calls. And like the AI will actually look at a, even at an Excel sheet, if that's kind of what the customer has, because typically connecting to the TMS has more nuanced and people working on MercuryGate or like Maclod is always trickier to like connect to first. But like we typically start with a simple Excel sheet and we just go from there. We give that to the AI and the AI can just go through all the numbers, all the loads in a list on an Excel sheet and just make an outbound call to that. And yeah, then the same with a dispatch agent at a broker, sorry, at a carrier. Just basically routing calls to the right department is the one that we always start with, with a with carriers. They just wanna have like a very personalized approach to like their customer service. Like instead of having like a press 1, press 2, press 3, which is annoying as hell, you'll just have like a nice human quote unquote voice that you can just talk to and you just say, hey, can I talk to James? And the AI will know that James maybe is in the dispatch department. So she'll transfer you to that department.

That's super interesting. Yeah, I used to have to do that in my old job working at a freight brokerage because people would typically hit zero and I was executive assistant. And so any calls that went unanswered, it would just immediately route to me and I would have to do this. So I immediately am already drawn to this, but I am curious. So you have these workflows set up and then you can kind of customize them from there. So what does, I guess, what does using the Happy Robot look like? Is it a Chrome extension? extension? Is it, you know, an integration into your TMS? Um, what does that— what does sitting— what is a dispatcher like sitting down at the computer? Are they actually like clicking buttons, or is this mainly like an automated system that after onboarding you don't really have to worry about?

Pretty much, pretty much. Once you have set up the assistant, uh, we have a dashboard where you can see all the calls that's happened, um, during the the last days and you can just see, okay, like this one call, we actually even show like tags. Like, okay, this one call was transferred to someone in sales and like we have a tag for like transfer to sales. I think you have that visibility. We're working even like on like dashboards to like just show you like what are the most common intents from people, right? Like, oh, maybe 80% of the people are just calling about load updates. We'd better build an AI to like give those load updates, which was actually the flow we've gone through with a with our carriers. Like they initially said, hey, I don't know what's going on with my phone system. I just want to have visibility into it. So what we do is like we integrate into their existing phone system, like be it 8x8, Dialpad, Vonage. They just transfer the calls to our AIs who sit on top of a phone number. We buy that phone number for the customer. So it's kind of that middleman that sits in between your existing phone system, very easy to integrate. And what we do is first build this router assistant who understands the intent in natural language from the user. Hey, I want to talk to Lizette in dispatch. And you'll just know that, oh, Liz works at dispatch, so I better transfer them to dispatch without knowing what extension Lizette works at or whatever, right? And then we'll have the visibility into What is the intent from people calling in? And then we can build like custom bots for the different pieces of the flow. As you saw there on the website, like we have already very established workflows for load updates. Pretty much the customer, the caller will say, "Hey, I'm looking for an update on load one three four five whatever." Sure, let me look into that. So we'll connect to the customer's TMS. We'll read. that load number, we'll look it up on the TMS, we'll see what's the status of that load and just display that to the AI so that the AI can actually just verbalize that in a conversational manner. It's that connection to the systems integrating to the customer's existing systems that is very powerful because obviously if you can actually look into what's going on in your TMS and provide actual information about the current status of a load, What's the usefulness of all these, right? That's a very big piece of our product, integrating into your existing systems.

And so you really start off with integrating into the existing systems and then almost start maybe the data collection, like phone routing first to learn the company. And then from there, I would imagine you would be able to add on additional functionality for that, I guess, sort of you know, Kate from the robot to handle it. Is that the way it works?

Pretty much. Pretty much. I can actually do another demo here with one of our other customers, Money Transport. Money Transport. If I call Money here, they're a carrier. We'll see that flow there, actually.

Hey, this is Kate with Moan Transport.

Hey, uh, I was trying to get a load update.

Sure, let me check on that for you. One sec. Okay, what is the load number?

Yeah, the load number is, um, 175 473.

All right, so that's 175473?

Yes, that is right.

Sure, one sec. It seems there might be a mix-up. Just to double-check, the load number you gave me was 175473?

You know what, I just made it up, so, um, thank you. If you can just transfer me to dispatch, that'd be great.

Okay, let me transfer you to dispatch. One sec.

So it's transferring to the actual dispatch department at Monnet. So basically what you saw there is me making up a load number, obviously, so it's, it's actually not gonna provide information to people just making up numbers, which is a big—

That's a good thing. It's a big— No hallucinating, I guess.

Exactly. That's the word hallucination is one that comes up a lot with AI and LLMs. That's why we build the systems. We call them proxy servers, kind of sitting between the AI and the customer's kind of TMS and systems that kind of cleans up the data and will only return information if you've been, in the case of Circle Logistics, for instance, if you've been validated as the right MC number, as a valid carrier. In the case of Monotransport, from the carrier side of things, we'll just validate if the load number actually exists, right? And we'll only provide a limited amount of information, right? We don't want to expose anything that the customer, in this case Monotransport, doesn't want to expose to other people calling in, right? Like, you don't want to give the exact coordinates from Samsara, which we actually connect to, right?

Oh, true.

Yeah, that's a very good point because I think there was an air— British Airways or British Airlines, I think they had an AI chatbot on the front end of their site and gave an inaccurate answer to someone that was chatting in trying to get a reservation changed. And it said it would be covered and then it wasn't covered. And then so the guy like ended up suing because they said your chatbot told me this. And he won the suit. And so I thought that that was a really interesting, um, just situation of, of how important, I guess, the, the reinforcement learning is and the feedback circle, the feedback loop. And I guess, how are you, once you onboard a new customer, how are you maybe, uh, protecting the, the intelligence of the robot? And so they're, they're only sharing the information you want them to share instead of I don't know, like how much revenue, you know, you made off of that load or something like that. How do you protect, I guess, your customer's information from maybe nefarious actors or accidental actors?

That's probably the biggest thing that I focus on these days because the AI backend, like the voice generation, the transcription, everything is set up now. The LLM part, the AI part is the one that I focus the most these days with the team. And that part of not disclosing information we don't want to disclose about the load or the payment or whatnot, that's kind of a very important piece. And how we do that is basically building this proxy server that I mentioned before, kind of like an intermediate layer that cleans up information before the AI sees it. We don't even let the AI see stuff that we don't want it to see, right? So any customer worrying about, oh, if I'm giving you access to like my TMS, it's going to just expose everything, right? No, we only have access to what you want us to have access to, right? Do you not want to expose the max buy rate of a load? Obviously you don't want to expose that. You just have the load port rate. You want to get the best load for the best offer for your load. So you're not going to be telling, or you don't want the AI to like be disclosing what's the, the max buy rate for that load. So obviously we're not going to be exposing that to the AI. Uh, she's just going to be able to like see a certain amount of information and like understand whether the offer is good or bad. Like kind of you're seeing how this proxy server we have like kind of serves as a, as a firewall to the AI.

Oh, that's super interesting. Okay, because you almost have to protect the AI from itself of being So that's a super interesting approach. I'm curious if, you know, the fine-tuning and the reinforcement learning, is it mostly the same for each customer or is it different for each customer?

It's—

the flow is the same. Obviously, the content of the conversations differs, but we have a platform now, like a part of the product is data collection. So like dataset collection. Every conversation I have with the AI, uh, we can create a training sample out of it and say, hey, this conversation didn't go as planned. I need to like, again, like I said, like very nicely reinforce it back into the, the AI, like retrain, like fine-tune again. You can use that sample that didn't go great, you can tweak it, you can just like refine the conversation, maybe obviously you're gonna— you want to refine what the assistant said, what the AI said, and you're gonna be using that sample as training material for the AI. So the same way that OpenAI, Mixtral, Anthropic, all these players train their models and fine-tune them for like specific chat use cases, we go a step further and fine-tune it to our enterprise customers for their specific, very nuanced and detailed use case. Like, oh, is it an inbound carrier sales rep? Then we'll gather data and training material for that use case and train a particular assistant just for that use case. And it's much more efficient, obviously, because you're not using like a very large LLM that is just generalistic and you're just like prompting it.

Hmm.

You're actually fine-tuning the assistant to like be good only at that. Obviously, it's not going to be able to help as a psychiatrist. It's not going to help you understand physics. It's not going to help you do math. It's only going to be good at sailing boats.

Which is arguably, I think, the biggest use case for AI is really those specialized approaches where it's not going to be, to your point, A lot of the large, you know, ChatGPT, Claude, things like that, they're going to, from what I understand, you're the expert here, but from what I understand, you know, they're going to reach a level of where they're all kind of working with the same information, but it's your data that's going to separate the usefulness of being able to take action or use it in a positive way that's going to benefit you. Is it, is that an accurate statement?

That's a perfect assumption. Yes, that's, that's totally right. That's what we've been seeing now with our use cases. There's been some use cases that we've been able to tackle just with an off-the-shelf GPT model, be it 3.5 or 4, whatever version. Unfortunately, Anthropic yet doesn't have function calling, but maybe soon. I say function calling, which is a key term here, because that's allowing AI to take action.

Hmm.

So function calling is a critical term, but I'll keep that parenthesis for later. To your point about like how we are fine-tuning for specific use cases, yeah, totally right. The more you can specify your use case and have an AI just tackle that, the more efficient in terms of even costs, it's gonna be more, it's gonna be lower cost. Because you're going to be able to fine-tune an open-source LLM on that specific use case. That's going to be cheaper than just bluntly using a GPT-4 for this particular easy use case, quote unquote. So the more specialized, the cheaper and the faster also. GPT-4 is known for being incredibly slow. So for real-time use cases like these phone calls, if you use GPT-4, it's going to be very, very painfully slow.

Mm-hmm.

Here you want latencies of 1 second, 1.5 seconds, 2-second stops. Obviously, if you're looking for information, you are— you're allowing the AI to like take a bit more time. Yeah, it's actually nice when like the AI is looking for the MC and it takes a bit more time because it just makes it feel more natural.

So that's interesting because you almost have to— you have to program that in there where it was, you know, I guess how much time does it typically take someone to manually check this? And then maybe there's a medium between I guess the robot doing it in a second versus the human manual, maybe you do it in 45 seconds versus a second. So it feels a little bit more natural. Is that, I guess that's your thought process and your approach?

It is. We just wanna make it sound and feel like a very natural experience. If you say you're MC and half a second later you have the verification, you don't even realize, like, sorry, what's— I was off the phone. I thought you were just gonna check it. I was just preparing my coffee here. Like, that's kind of what we wanna allow the callers feel like. We just wanna make it feel like a very natural experience, right? So in that sense, yeah, any— anything that we can replicate from the human side, we try to do that.

What about— how do you, I guess, from, um, like a philosophy standpoint, how do you preach the values of like Happy Robot or just AI in general to a lot of like the doomers out there? Because there's a lot of them and they're very loud.

Yeah, and I understand that. I mean, it's, um, something that you don't maybe fully know about, it scares you, right? I have to be clear, like, AI is just statistics. It's no magic. It's just, uh, a bunch of statistical predictions. Like, it's just predicting what is the most probable next word, uh, that I should be like saying next. So like AI is kind of dumb if you think about it. Obviously I'm just like making it feel like very stupid, but I mean it's super smart. But it's smart because we— it's in, it's in a lot of, um, a lot of data. But under the hood it's just statistics, right? So nothing to worry about unless we use it the wrong way. It's kind of like a knife. Like, you can obviously kill people with a knife.

True. Very true.

But you can also cut tomatoes and prepare awesome meals, right? So it's everything in life. It's just a tool. It just depends on how we use it. I don't think we should ban AI for it being AI. It has its uses. We just need to know how to leverage them the right way. And in logistics, we've been hearing a lot of those naysayers, especially smaller brokers maybe, in some of the LinkedIn posts that I did, like saying, hey, this is never gonna work. We need more humans. We need more warmth. We need more conversations. And I totally get that. But when it comes to like really transactional phone calls that some of the largest players have, It's actually great to be able to like offload your team off of those transactional calls and just get them to actually build those real connections. Like who wants to give payment status updates? So like a factoring company, who wants to just say, sorry, we cannot work with you. You don't have enough months, like your carrier, you know, like who wants to like invite, who is saying no to someone over the phone? It's painful. But if an AI does that, It's just business, right? And unfortunately, that in this case on the carrier sales use case, kind of diving deeper into that, if a carrier cannot work with Circle or whomever, they just cannot work with them. Maybe they have some penalties that they got and they don't have enough months that they've been working for. So Circle would just immediately invalidate them. Why would you have a person saying that to Do they care? You just have an AI gracefully and very politely saying, sorry, we can't work with you guys. If yes, yeah, let me transfer you. So it's kind of the, again, like going back to how you use it, right? How you use AI, it just depends on the people using it, right? There's gonna be people that use it for bad stuff. I mean, I've already seen, especially with like voice cloning, which I can actually do another quick demo now with how I talk to myself. I cloned my voice, so I'm able to like talk to myself now. People are going to use that to like just do bad scamming stuff. I mean, that's already been happening. Do we now stop any AI initiatives because some people, some bad people are doing that? I would say no. I would say we go for— we go after those bad people, bad players, and we try to reprimand them, right?

And Yeah, I, I think that that's, you know, it's very much the mentality that we've seen with any kind of technological advancement. You know, there, there are people, you know, that were very upset when, you know, Henry Ford gave them a car instead of horses, when, you know, farmers got access to a tractor instead of, you know, using human labor. You know, the, the role of like a social media manager didn't really exist, you know, 10, 15 years ago, but as new tech enters our society and enters our, you know, workflow, then it's just one of those things that you can adopt it where it makes the most sense for you. But jobs evolve, jobs change. And to your point, the, the transactional calls— is it something that, you know, drives people to want to go to work in the morning? They want to build relationships, they want to make money, they, you know— and I would, I would hope that a lot of people want to do it in an ethical way. And I think I'm a firm believer of using these types of tools like Happy Robot to help companies grow in a way that makes the most sense for them. I mean, AI adoption for my own company has helped tremendously where we're able to publish more content than some of the biggest companies in the world and at a frequency that not many big teams can match. But we're small, we're agile, and we have our processes mostly dialed in. But then from there, you can really utilize your team to focus on the things that, that needs the human touch. And so, you know, having those, those conversations, I think, you know, routing, you know, calls, if it's an escalated manner or matter to, you know, somebody that can— is a, you know, real human, I think is probably the best use of their time.

100%. That's exactly what I, what I try to Explain when people have concerns about AI, it's just getting people, giving people more time back. Like give your team more time back to actually do the actual sales calls, to actually build the connections with the carriers you actually care about. Like those type of conversations are the one you want to have your team working on.

100%.

Anything that is out of that realm, yeah, you can just use AI or Just let's call it like tools. I mean, it's just tools. Like AI is just another tool. So aren't we using email, some email, like before AI, quote unquote, and before AI actually became like known as AI in the common space, before we had AI, we had email automation with, you know, other flavors. So is that also bad? Like—

Spell check.

Spell check.

Google search autofill. Like all of these different things are a form of artificial intelligence. But, you know, for whatever reason, people have chosen this to, to be, I guess, their tool of fear. You know, that, you know, try to restrict people from advancing in technological ways. And, you know, I, I'm, I'm on the tech side of things, especially, um, from the mindset of, uh, that— I don't know if you've heard of the people, I'm one of them, that, you know, is very polite to like ChatGPT just in case it turns on us and they'll remember who the polite ones are.

That's so funny. I, I've heard that, and I, I tend not to do that myself. I, I'm now more like a, like, super direct, like, do that, do these. I sometimes find, like, am I doing something wrong here? Like, will, will it— will this come back to me at some point?

Yeah, that's definitely a, a funny approach to it, is, you know, saying please and saying thank you just in case.

Just in case. Just in case, you know. No, this is, this is fascinating. I mean, I think Also, to the point about like people maybe losing jobs or whatever, I've never heard— no one yet. I still have to hear someone tell me, oh, I'm gonna fire people because of these. They just put them to other tasks because they want to grow. I mean, companies want to grow, and you're not gonna just get rid of the good people. Obviously, you want good professionals and employees to like stay with you but do more value-added tasks. So I've yet to see someone saying, I'm going to just fire people because of like this cool AI thingy automating some of the tech calls or payment status updates or whatever. They just put them on other more value-added tasks. That's kind of my learning and my reading from these.

So what do you have a, maybe I imagine you probably have like a product roadmap. First of all, how difficult is it to have a product roadmap when so much changes in the world of AI from like day to day?

We just almost don't have it. Like, we have it like flexibly on our minds. Like, we just keep talking about it all the time. But, uh, it's crazy. I mean, every day you have something new, and, um, you cannot— you cannot keep it up, keep up with it. During my time at the PhD, you already felt like in 2021 or so, you already felt that like there would be like a new paper every week. Now it's almost like every day there's something new. It's like Well, I mean, great, because we're getting better tools even for ourselves. The better LLMs we have from other providers, the easier the prompting and the fine-tuning is for our specific use cases. So that's great. But it's still hard to mentally just take some time off. You're just constantly thinking about what's new and how to improve it. But yeah, I mean, our roadmap right now is just building a very easy-to-use platform for any logistics player to just like build different use cases on their own even. Obviously, when the use cases are complex, we come in with our engineering team and solutions engineers and we like tweak the use cases and fine-tune those use cases with real-life data that we gather with them. But for like easier use cases, like, you know, like that router AI assistant, That's kind of plug and play. Customer comes in, I want to automate my phone tree, I want to do some load updates, those packages are kind of built, pre-built. So it's kind of that level of self-service that we want to have in the platform more and more so that anyone can just build cool, quote unquote, easy AIs. And then for like the real hard use cases, currently today, the level of fine-tuning you need is still there. Like, it's not as easy as just like prompting the same way that you prompt ChatGPT where you're like talking to it. Prompting gets you so far, it's only so far. So typically we need to sit down with like, uh, big enterprise customers like Circle, starting to work with other, uh, from the top 50, uh, US brokers, like defining their nuanced use cases. Then again, they typically look the same all the time. Like, it's just carrier sales, it's just payment status updates. The only difference is the connection to their existing systems, right, and TMSs. But we're working on that, connecting to other players in the industry to like have like this environment that these kind of— yeah, this environment of like tools in logistics that we can just connect to and like serve anyone through any tool.

So what does sort of the ideal customer look like for Happy Robot? Is it, you know, at the enterprise level right now, mid-market? You know, what is your ideal customer that your solution is built for right now?

We're very comfortable in the enterprise arena, just having resources available also from the company, bringing their own engineers, They know how to operate their internal systems.

Very true.

The bottleneck we've seen so far with more mid-market folks, I guess, would be they rely on other TMS systems and some TMSs, they take longer to let you integrate to them. So as newcomers, we're still battling that in a sense. Can we connect to America Brigade today? Takes some time. McLeod takes some time. So like some customers that already have their own TMSs, uh, or even like, um, have TMSs that are more API-friendly, that's always, uh, kind of the, the better customer for us. We always overcome that through different approaches. Like there's also web scraping or like app scraping, so you can have like a bot. We've had some good success with a um, ask an app scraper that would just look at the TMS and give us like the information every, every 30 minutes. It would just like upload it versus having to wait for the TMS to give us access to their APIs, which typically they don't love for some reason, which I don't understand as some engineer. Like, being an engineer myself, like, I'm like, why?

Why?

Like, we have a document, we have like open APIs now that people can just use. Like, we also have that product. So to your question as well, like we do shine with enterprises. We also have our developer kind of tier, if you will, where other startups in logistics are actually building on top of. We have some startups building products for factoring companies. So they'll build a bot that in the name of a factoring company calls in a broker and says, hey, calling about this load number. Uh, is this associated to like carrier, um, ACME Transport? Yes. Okay. Uh, any, any advancements? Uh, no. Okay, all good. Thank you. Bye. So we have some startups doing that with our APIs. That's why like I'm sad to hear that other players in the industry are like almost closing the door in front of you so that you cannot even like integrate to them. But hey, I think, I think things are getting better. some of the new TMSs are already like OpenAPI friendly. And that's, that's great to hear.

Yeah, definitely. Especially for a supporter of this show, Tye Software, Tye TMS. I know we just had a conversation with them last week about how they prioritize integrations and having that API capability. And just, it feels like there's been a shift over the last like 3 to 5 years within TMS platforms that they realize that they need to be a little bit more open. I know we have taken a lot of brute strength to get here, a lot of convincing. Some of them are a little bit slower than others to adopt the, you know, new technology, and to each their own. But it— the advancements when it comes to how much more productive that you can be throughout the day if your platforms can talk to each other is just astronomical, you know, and it really does allow you the ability to put people in their best position where they're going to succeed internally. And then ultimately, hopefully, the company succeeds, customers succeed, you know, it's a full loop circle. All right, Pablo, is there anything that you feel is important to mention that we haven't already talked about?

I would say, don't fear AI. It's just going to be a tool. Know how to use, learn how to use it, learn where it can be helpful. I've seen a lot of people being very successful on AI for email as well. We don't do email. We actually decided explicitly not to like do any email stuff. So if anyone wants to like start with email, I have a good number of friends that I can refer them to doing some cool AI automation through email. AI is gonna be there, so you'd better learn how to use it and acknowledge it.

Acknowledge it.

Yes. So if you do wanna implement, like if you're a broker out there, a carrier out there looking to automate some of those transactional phone calls, they can reach out to me at founders@happyrobot. I guess that's a good email, founders@happyrobot.ai. They can just reach out to us there. Just visit happyrobot.ai and yeah, we'll Follow us on LinkedIn. We do have a fair amount of good posts that got some attention, good and bad, I should say. You're gonna see a lot of people just bashing us and others saying like, this is cool. But hey, that's kind of what every new technology suffers from at the beginning, I think. So excited to be working on these and thank you for having us, by the way. Really appreciate it.

Absolutely. You know, I guess to the haters' comment, you know, you're not doing— you're doing something right. If you have a few haters out there that are skeptical, but obviously you have paying customers and you have a bunch of people around you that believe in you and believe in your mission. And you're proving it because you have actual customers that are paying for your product. So I think it's really cool to see. And so kudos to you and the team. Where can folks, I guess, check out the Happy Robot website? We'll put that link in the comments. They can get signed up for a demo and hopefully it's a good fit for you guys.

Exactly. Yeah, happy to, happy to do that.

Awesome. Thank you so much, Pablo.

Thank you so much.