Building the AWS of AI Work: HappyRobot CEO Pablo Palafox on Deploying AI Agents Across Enterprise

0:00 Our take is agents will build the UIs and dashboards and evals and exactly and databases and to show the work they're doing >> that in a way that matters to the to that customer. We have like so much going on that that we're like so excited about style like we get very excited about stuff and and we we try to like never lose focus but >> uh it's it's not easy sometimes. This is I literally bring my my wife who is a teacher >> and has vacation during summer to label like data sets. The company right now is it is it's crazy. It's nuts. It's a lot of work. Um you know like the bar is super high. But there's that sense of like >> we're all like having fun and like like laughing a lot by >> building some agent that actually messes up uh w with with like some stuff and he's like oh [ __ ] Like that's crazy.

0:50 >> Cool. All right. Welcome back to another episode here of Skywatch Blue Jay's uh carb podcast. Today I got a very special guest. I got Pablo from Happy Robot. >> So happy to be here. >> Yeah, absolutely. Why don't you introduce yourself? >> Hey everyone. Uh my name is Pablo Palafog. I'm one of the co-founders and CEO of Happy Robot. Uh we don't do physical robots. Uh we build agents for the enterprise. We're building the agentic infrastructure for the enterprise to build their their own AI workforce. We can talk more about Now >> do a lot of people get confused that you build robots like when you introduce the first very common thing is like >> it's actually uh it's actually like my intro line now like we don't do physical robots >> right >> but I need to make it like just very clear >> we literally had people in some conferences come to us and say hey I saw happy robots so like >> uh we have a warehouse where you know like we we need robots and like well >> sorry we don't help you with any >> disappoint you but it's actually >> but you guys have stuck with the name still We're stuck with the name. We We like it a lot. It's like >> it's different. People People remember it really well. Yeah.

1:52 >> And what we like is that people don't mess mess up the spelling as much as other names. >> Uh people still add like a little u empty space like an empty space between it's fine like that that I can take. But like >> uh in the beginning we were joking about fancy names like get messed up all the time. Yeah. Did you did you have other names in mind that you >> The thing is when we started and I I've shared this in a few other podcasts maybe. Um we we started doing computer vision.

2:18 >> Okay. >> So we started doing computer vision. I was doing my PhD in in deep learning and and computer vision in Munich. >> Yeah. >> Uh so the three founders in in Hamp were Spanish, right? >> Uh one is my my brother and the other one is my best buddy from college. Yeah. So we we go a long way. Uh I actually have known him more than than for longer than my wife. >> Oh, really? Fun fact. Uh um but yeah, my my buddy Louise and I we started doing, you know, small projects at in uni. We we built a an under robot underwater robot, like a tiny submarine >> uh in college. And then from there, we went and worked at the same startup for killing drones. Basically, we we needed to like identify and like like shoot drones literally for the Spanish government.

3:01 >> So, you've been in hardware? >> I I was in hardware. >> I was in hardware. Yeah. Okay. >> And then we end up we both end up in Munich uh for our masters and then I stay for the PhD >> uh and after an internship in in Meta uh at the time Facebook >> I'm like like I I need to get out of this like slow moving no more like I mean at Meta I was already doing like computer vision and deep learning. So like no hardware really. I was using some camera setups and like you know like uh I was involved in the um right like autonomous driving team in in the university there in Munich >> but it was more around the like building you know uh some some neural nets I was using some pietors at the time I'm like >> uh >> clueless these days if you put me on like actually training some some some of that stuff Louis is still like so involved there but >> um yeah we we end up we end up saying okay I'm going to drop out of the PhD Louis is going to quit his work at uh HPE.

3:56 >> Uh he was working on some cloud computing stuff at the time >> and Kabi he was the CFO of this uh distributor this olive oil distributor. He was in logistics at the time. >> Okay. >> Uh and we we all three quit apply to YC get in summer 23 and >> okay >> uh start building uh some computer vision platforms. >> Uh >> so it was never related to voice or >> it was logistics >> or was it still in supply chain? It never it didn't start as as that like the the the v 0ero of hyper robot. You might as well call it like a uh something else.

4:24 >> Uh it was around >> building a platform for for robotics companies to actually build computer vision systems. >> And we said okay we're we're helping these robotics companies uh >> right >> really train their vision systems. So we're making their robots happy >> because they can see now. >> Okay. >> All right. Okay. Now it makes sense. Okay. >> Now it makes sense. Right. And then when we started when we pivoted after YC and I've told these story a few times when we pivoted after YC uh we started building around uh some of the LM stuff that was happening. This was >> September 2023.

4:59 >> Okay. >> So we started building some some stuff and we ended up building a voice agent. >> Uh at the time there was not a lot of stuff going on. >> Uh I think it was our friends from from blank. Uh I think By was starting out. Uh so it was just a few of us building building voice agents and we >> decided to take a bit of a of a weird approach of an interesting approach in that >> we built a lot of it inhouse or we we literally built our own like voice architecture inhouse like it was essentially you know competing with our friends from Vappy and and really at the time uh >> and then we decided to specialize in an industry uh that that we that we knew that was uh through habi uh through one of my co-founders that was logistics >> right And we started there and we started digging deeper uh into that into that space.

5:45 >> Yeah. >> But we always built a horizontal platform. >> Uh so we always had that in mind. >> Okay. So you always you always had the horizontal platform in the back of your mind. >> We always had that uh as as part of the product because we needed to expedite the way we deployed agents with our own customers. Even within logistics and and supply chain, there's a lot of different workflows that you have to build, a lot of different types of agents. And if you think about it, an enterprise in logistics or supply chain is not that different from an enterprise in the telco space or in the banking space. I always say it's all a combination of people, processes and data, >> right?

6:21 >> Uh and if you can build software to really, >> you know, like automate processes in in a in a big enterprise and supply chain, you build an agent or you have built an agent or a system that can automate processes in backing, right? Yeah. >> So are you saying then that there's less and less value around being vertically specific? >> The only value is in speaking the customer's language. >> Okay. >> There's less and less value in the productization per se.

6:47 >> What about like the the like the actual end to end experience that you end up giving to users? Our take is agents will build the UIs and dashboards and evaluating things that are >> exactly and databases and and and systems of record. >> Yeah. >> To show the work they're doing. >> Uhhuh. >> That in a way that matters to the to that customer. So even you know put a put the example of uh DHL or FedEx. No, even even there >> the processes that they each have they're different. uh and we couldn't really build something we it's impossible or it's it's not the right choice today to build software that is static >> in a way that uh >> tries to cater to both these this this supposedly same type of industry companies right >> uh so we we're very much believing in the specialization of in the personalization of software right >> I kind of started calling it on demand uh UIs or on demand >> on demand dashboards um >> because agents are doing that and they need to surface the the issues and the instances where they need human intervention. Uh so they're now starting we're building software to basically let the agents build their own UIs and dashboards. So >> Gotcha.

8:00 >> We started again with that horizontal agent builder. Now we're going more into the into the data piece and combining that with the with the UIs and dashboards and >> Gotcha. >> So you guys are very much moving then beyond the supply chain logistics space. >> Exactly. We are taking a bit of a multi-vertical approach. uh super super focused on on our core industries and and that just the first one just happens to be supply chain. Uh we're going to continue investing and and overinvesting there because there's a lot to automate.

8:29 >> Uh but now we're kind of getting deeper into other verticals already partnering and and working with some folks in Telos uh utilities uh financial services. Uh so so yeah again uh ultimately the the processes that an enterprise has they're all so So similar in a way. >> Okay. Interesting. Okay. And what what scale do you guys currently operate at at least in the supply chain logistics and maybe just in general as well for everyone to hear?

8:55 >> I think we're at around maybe half a million half a million to a million like runs a day. >> Okay. >> Runs can involve uh >> a phone call with a follow-up email or like some form of like reasoning, right? >> Um it's roughly around that. Okay. >> I'm going to say >> that's very impressive. That's insane. >> It's uh it's getting crazy. Yeah, it's getting crazy. We have over a 100 enterprises. Uh >> that's awesome.

9:22 >> Some that are maybe more public about it like like DHL in logistics now. We're partnering with um >> um one of the largest banks in in in Spain uh and in the UK starting to do some some cool stuff in France. We work with uh one of the major um ocean carriers in the space of you know uh logistics deal if you will. Uhhuh. >> Um, yeah. No, it's really >> that's super exciting, man. Congratulations. That's awesome. And now we're opening a bit of a of a detail natives kind of division where this is like still a bit uh a bit uh in the making and we haven't shared too much but >> um we we started seeing some startups that were building on Hamper Robot mostly because we let them in as friends and then they went bananas and they started like uh automating a lot of their own stuff with with Hamper Robot as the infrastructure for them to build agents on. Okay. Um yeah, which is >> so that's almost becoming like as you mentioned like a vappy or a bland in a sense >> less developer maybe uh or how to put it less of an open like like super open to like any developer more like >> are you maybe like a series D series you know like series D plus start with like already like a decent you know like um spend if you will on >> a sales team a customer service team um a payment collection team and we we have this one this one uh this one customer who's >> basic basically building uh agents for recruiting.

10:51 >> Okay. >> Um and they've built on a lot. Okay. Interesting. >> They're like a I don't know they've been around for like 10 years. Uh >> several like um I don't know if like uh >> dozens of of millions of revenue and and probably more. Okay. >> Uh so pretty stable company. Like we we we overinvest in those. we we kind of help them with um kind of thinking about the architecture of the agents and like we put a lot of effort into those partnerships versus more of a super entry developer.

11:22 >> So how does it how are you guys scaling this model then? Right? Because I'm guessing like >> all of these companies are going to require a lot of support, a lot of effort to really make it work in production, right? At least from my experience, it's not as easy to get them working at like high automation rates um where the customer is also satisfied. >> I'm you guys have a huge FD motion. Is that correct?

11:46 >> We we do. Okay. >> We do. >> And talk me through that like how how does your FDE model work? Are you bullish on that being kind of the way to kind of grow these companies going forward or do you think it'll be something else? >> It's really a mix of over or investing in more automation and more product. Okay. >> Uh we now have our meta builder um which is really a >> uh a way for a user to build their own agent just by using natural language you know uh nothing groundbreaking but really expedites the process of building these agents uh amazingly.

12:20 >> Uh so you just say hey I need to build like a like a like an agent for customer service. >> Uh oh fix the problem here. Like maybe even in production already you you say oh like fix the problem here. Um I don't know what happened. Um, >> yeah. >> So, we're over investing in that in that motion. Oh my goodness. Someone someone had a bit of an accident there. >> Uh, my goodness. >> We'll we'll cut this part from the >> Yeah.

12:45 >> Uh, but um, so yeah, we we we see the we see the framework. We we've basically built a bit of a framework around um this self-fixing agent, >> right? >> We call it BEA or >> Okay. >> Bea Ba. uh build a valid audit. So essentially >> the build piece is >> helping you metab build your agent, right? You can very quickly build your agent uh by by using natural language, >> right? >> Before we go into production though, >> you run a set of of evals, >> right?

13:18 >> Uh and and you basically are sure or certain that at least your current set of evals are passing, right? >> Uh with the ex with with that with that status with with that stage of your agent, right? Then you go into production, >> right? >> Um we we actually have three environments like production, staging and development. So you can play with that. But let's say we've already promoted the the use case to to production. Maybe let's say it's a it's a sales agent >> um that uh is calling, you know, dormant accounts uh in batches of 10,000 calls a day and it's following up with an email and it's capturing leads for the customer. Right.

13:52 >> Right. say that now you're in production and out of those 10,000 you know like uh calls a day >> you obviously cannot audit all of them at once with or with with a person you can go like one by one >> so now enters the the auditing piece right >> so that's what we call the the the AI auditor uh you can enable that uh in the in the product today and >> just basically set a a percentage of calls of of runs of executions I'm just going to be phone calls for simplicity but it can be like anything you know we do a WhatsApp email a lot of lot of um SMS but um let's say if it's it's those 10,000 calls so you enable that auditing to basically flag potential issues bas based on north stars >> right >> those north stars are automatically generated from the prompt or based on your own like guidelines right >> and now we close the loop because if you've identified a call that went poorly maybe you know like uh the northstar was always wrap up the call cross-selling on another product and then in that one case like the agent didn't do that because your main prompt didn't follow that for whatever reason >> now you can use the meta prompter again >> to to to fix your agent based off of that uh >> AI auditor failure uh that failure the it's that loop right so that is that is what traditionally the the the FDS were doing right >> um >> and when I say it's a mix is because the edges are still super involved in in scoping and and kind of digging deeper into these massive enterprises. No, we just had an on-site with >> with one of our customers uh one of our prospects really these the last week and we had a team of four people going there. It's it's one of the major um uh one of the a fortune 100 uh enterprises and and >> the the interesting thing is the hardest piece is not like technology is actually change management and understanding like oh who's going to get pissed because we're going to build this customer service agent and they actually fail doing that internally as their as their engineering team. So we have to like uh be a bit of a ninja there in how to navigate these orex kind of catering to both business and um uh the technologists the the engineers >> because we do want to be an extension of their engineering team right >> and we honestly believe that that that that is possible because we have seen engineers and customers building on HA robot and like getting a lot of um um praise from their uh colleagues >> but also you know like there's that sense of I could build this myself.

16:23 Yeah. >> Um, so, so we also want to go around that a little bit and go to the business people and get their >> Yeah. No, I think you're Yeah. I think there's always the inclination of an engineer to want to do things in house >> versus versus hey, like can you use something else? So, it makes makes a lot of sense. >> You mentioned something. There's like a like you like navigating this Fortune 100. It seems like there's different personas here. There's like your business side of it >> and also the technical side of the persona. What are you typically seeing when it comes to adoption and like how are you tackling it?

16:57 >> It's interesting. It changes a lot from company to company. Like I've seen some innovation teams are >> um highly motivated to use something like like >> like this technology like pro like an AI agent uh builder really uh to partner with and move faster. Yeah. >> Uh and they almost bring it to business like guys like we found the thing that can help us go really fast. Um, in other cases, we did have to go to ultimately the CEO and say, >> "Okay, this is a seven figure partnership. Uh, this is what it takes to partner with HRO, but like help us out." And then it comes from like the CEO and it pushes it down, you know, to the to the team.

17:33 >> Uh, so I've seen both motions. Um, >> I mean, ultimately I I think we're going to see more of the engineering side, meaning we have to go through the CIO, >> right? um or the CTO. Yeah. Because we're seeing more and more that CIO, CTO, CPO, right, >> are like the new >> um almost like the new HR department. >> Yeah. >> They're hiring agents to do the work. >> Uh and they want to make sure that whatever is being built in their companies >> is rock solid.

18:02 >> Yeah. >> So that there's no like, >> you know, problems downstream or or down the line or like, oh, like an agent like uh we we partnered with the wrong company or we used the wrong thing or whatever. and these agents just leaked all of our information because it was architected wrongly or whatever, >> right? Yeah. Yeah, makes sense. So, we're >> And what are the deployments like looking like in terms of scale? Like are people like auto deploying these things into production at like 100% volume? Is it still like scaled down at like 10%, 20%, like what do the rollouts look like?

18:35 >> We've seen rollouts that have gone to like 100% like usage in a matter of weeks. Okay. It >> totally depends on >> is this across all modalities or very specific modalities? >> For whatever reason, voice for us has moved the fastest. Um >> email is maybe still like a modality. You ask modality for >> like Yeah. Yeah. Text. Yeah. Voice. Uh email. voice has maybe because we have like a really solid uh we you know we build our own voice agent to the extent that we even have like um our own text of speech actually for a few different languages now which is >> which we haven't talked too much about it but uh maybe you know in a few weeks we'll we'll share more >> uh but small snippet there like we we obviously you know still partner with some of the uh major like Texas speech writers but um more and more like we we see like a tendency to to own the commodity uh ultimately like we when when we deploy on on when we want to now deploy on prem with more customers and we have done a few uh not really on prem but like on their own VPC and things like that like it's very helpful if you >> like have more control over your your tech stack.

19:43 >> Um so on the voice piece let's maybe double down on that. Um >> we've seen customers go from a week of testing let's say week number two after scoping it out >> to 40,000 calls uh per day >> right? uh this was uh this was actually with with with u with DHL actually uh and we've talked about these use case publicly so I can share a little bit more basically it's calling uh consumers >> in the US >> uh to let them know that they have duties unpaid >> okay >> you're literally calling consumers which is traditionally one of the hardest things like >> and and you're just letting them know you know you have this package you haven't paid the duties and then typically what happens is the person doesn't really know that they had it so they start asking questions and then the agent can navigate those like yeah I can explain like this is the the the shipment number and by the way like this is how you can pay or do you want me to help you pay it? Okay, let me guide you through the process on the website.

20:36 >> Yeah. >> Uh and what we saw is like a massive conversion like a massive increase in conversion for for deal paid. Uh so this is the scale at which we can go now like we can go from like a few you know 100 calls per day when testing uh that was maybe you know I was saying like week two uh like week four of that deployment we were like already at a at a massive scale >> 30 to 50,000 calls a day or something like that and it continues to be to this day. So these sort of of scale we we've seen and >> have you guys gotten to the point where you're building like intell like intelligent operations of these agents. So like if I start on voice but then a customer moves to chat or email like it's carrying over that context.

21:16 >> Exactly. >> Uh we we kind of name that uh contact intelligence in the in the product. >> So essentially you >> you have like different modalities that you've built your agent for. Say that you have uh WhatsApp uh and and phone uh and and email. We kind of see that combination uh relatively often, >> right? >> Um interesting. We have like the three uh autonomous driving cars here. >> Yeah, we do. who's going to go first.

21:42 So you have um you have those three modalities and then you kind of enable contact intelligence, >> right, >> for for those three agents. >> Yeah. >> H and they all share the context uh from you know that phone number they called in. Yeah. >> Then now we realize it's actually associated to this other email. So you can start like auto injecting context from the conversation. So that's picking up a lot and I mean it makes sense, right? Like ultimately >> when a customer when we say that >> maybe point solutions are less >> um helpful for in the long run >> uh is because that sales agent >> uh that found out something about the specific customer >> that information can be used down the stream for like the customer service agent or the other way around. We're working with some Telos uh where the customer service agent can actually cross-ell uh the the the customer after fixing a problem with the router or something, right?

22:35 >> It can crossell it on like oh sounds like your Wi-Fi is actually you could you could be paying a little bit more but like for better Wi-Fi. Do do you want to explore that? >> No, it makes a lot of sense. Yeah. I I think like even in like so before Blue Jay I was thinking a lot about just voice agents across like a bunch of like customer interviews for example and then also even like very like restaurant industry in both of those use cases right for customer interviews that can directly feed back into product decisions >> in restaurants like if I realize that someone is ordering the same thing every time the next time they come back I should offer them that maybe I say based on their preferences I say hey why don't you try this right and so that type of intelligence is always really interesting to Y >> and so you guys are now able to do that.

23:16 >> Exactly. >> Okay. And it doesn't matter what modality doesn't matter. That's freaking cool. >> Okay, that's awesome. That's exciting. And what what has been like from an engineering perspective, how have you guys built that memory layer? Are you guys using any tools? Is this kind of built purely in house? Like how are you guys thinking about it? >> Really purely in house. We we take a bit of um >> of a of an in-house uh kind of build approach, right? uh at the end you know you know Lou is a little bit uh >> he's a little bit obsessed on on control in a good way not like controlling the the tech stack because it gives us a lot of flexibility >> um and this is actually you know I was not >> on board initially like when we started 2 years ago I was like man like we're going to move slower because like you need to build a lot of these stuff but now I I I >> it's paid off >> I it's paid off so much like I >> I totally >> like take it back and be I was wrong like he he was so right like just owning the the the textile really helps you move very fast um in the long run, >> right?

24:14 >> It does take a little longer but >> right because you're able to own a lot of those decisions deploy is that the idea? >> Exactly. Now, now for example, we need to um you know fix maybe some of the uh some some of the way like we extract those memory snippets from from the conversations in WhatsApp or in email or whatever. Yeah. Yeah, >> you can just go to uh Danny, one of the engineers, and he'll he'll fix it right away, and he knows where to tweak things, >> right?

24:40 >> Okay. Super cool. That's awesome, man. >> And so, what's the you guys now have I guess like your your hands in a lot of different places, right? There's every single modality. There's like these multiple industries. Where is Happy Robot going? It's like there's people now using your platform to build on top of it. You're deploying things. You even maybe have your own TTS model. >> Where where are you guys going? Like where is this what is the end result here?

25:07 >> We our goal is to help enterprises reimagine the way they do work. >> Okay. >> We want to be that uh that AWS of AI work. >> Okay. where you know we we might not always um build like we probably will not build like our own LLMs although you know in the past we actually had to know like right >> uh >> I think that speaks to the fact that >> you'll we'll do whatever it takes to deploy agents that work we'll do whatever it takes to deploy agents that kind of get things done for our customers >> uh if you know two years ago it was fine-tuning uh llama and mistral and we literally had to do that because GPD 3.5 at the time was was not great and and GPD4 it was super slow, >> right?

25:46 >> That's the way we operate now. Like we'll do whatever it takes to be that orchestration layer >> for the enterprise to deploy AI workers at scale reliably and in a way that they can keep track of of things they're doing. >> Yeah. >> And make sure that >> they have peace of mind. No, like they have peace of mind. >> This company Hero is helping us deploy these AI workforce. Um it's super transparent. We see like what the agents are doing. We we can modify them. It's not a black box in that >> we don't see like the prompt that HRO is using for evals or for uh the the prompt the memory snippet extraction. It's all super transparent. Like if you go on Hypro like you'll see that it's >> it's all exposed. We don't hide any prompts. We don't we don't have really any like underlying >> you know logic if you will that is abstracted away from the enterprise.

26:33 >> Oh so you you open source everything to your enterprise. >> It's visible for them. They can tweak the promps. They can can see all of that. I mean there's some things for example for for the for the agent now we for the text to speech for example right >> um >> um we h we can like make it cuff a little bit or or laugh it's still like in beta uh I think in 11 labs like had something really cool which is u similar um >> um >> why did you guys do your own model >> like what was the idea behind that >> it gives you flexibility again so for example and we we still we still know like partner closely with with these folks and and we will continue >> um >> but in reading numbers for example um we had a few issues with with uh with some of these text to speech providers >> and and for us reading numbers out loud uh is super critical.

27:25 >> Um so what we decided is okay we we need to at least have the option to you know have an agent that can properly pronounce >> a shipment number like 1 2 3 4 5 in a very clear manner. >> Yeah. uh in some of the other versions out there like the agent would just like say something one of the what do you say like you said again so we struggle a lot with that so >> it goes a little bit of the to the point of the more control you have over your agents the more enterprise you can go >> um because they want to they want to be closer to the bare metal in a sense >> so that was a bit of a decision but to your point like yeah now going like different verticals um having like different parts of the product.

28:09 >> We we don't really see it as like >> it's not really that hard as in >> the let's start with the multi-vertical piece. It's all about speaking the customer's language. And >> again, it just takes very smart people >> to sit down with a customer, understand the industry for for, you know, a few weeks even. >> Once you've deployed our first your first customer in an industry, now you know a lot about that industry. >> And that's what would happen to us. No, like when we started in in supply chain or in freight brokerage even which is a part of supply chain.

28:38 >> Yeah. uh I was deployed in some customers for a few weeks or a few months >> and I ended up like literally doing >> the work that they could do you know >> uh so it just takes a little bit of care and attention to detail >> uh and really trying to fix that customer's problems right >> so that that is more of a go to market thing from a product perspective like there's nothing really that blocks us from >> you know doing >> being able to do all of these things >> exactly from like the agent being able to read documents and and reconcile invoices >> to to browsing some some websites. Uh and obviously there like we we haven't built our own browser agents but like we we do use some some amazing vendors out there. I would actually have to check who but >> we we do rely as an orchestration player as an orchestration layer on on existing capabilities out there.

29:29 >> We'll always build whatever is not there yet. So not we're not we're not stupid. you know, like if if the technology gets to a point in in whatever modality, whatever, you know, aspect of the product, if it gets to a point where we don't need to build in house, right, >> we'll just like integrate that into the product. Ultimately, we're here to deliver agents in the enterprise that do work, right, >> from sales through customer service uh onto uh collecting money and and reconciling invoices, right? So, >> no, it's a it's a it's a cool very big ambitious ambitious. Yeah.

30:01 >> It's ambitious. Yeah. And what is um is the goal here to just become like intelligence layer for for these companies and enterprises? Is it more to be an infrastructure person for these enterprises and kind of let them generate intelligence in their own way? Is it a bit of both >> or something else that I'm maybe not? >> No, I mean, you know, infrastructure is probably a good way to define it as in I'm giving you the the the the tools to really define how you build your agents.

30:34 >> And the enterprise really appreciates that flexibility because >> we understand we're not going to be the only, you know, agent >> infrastructure for for a, you know, multi-billion dollar enterprise. they'll be using all their systems and they actually ask us ask us a lot about oh do you support A2A frameworks or like MCP blah blah blah. So we understand like we're going to have to like >> like be nimble and and adjust and and be helpful where needed, >> right?

30:58 >> Um >> but obviously the intelligence play or or bringing the the insights from the work the agents are doing >> Yeah. >> is really key. like the fact that you can gather so much information from these, you know, from from this work, >> from these conversations, from from these interactions with customers, with partners, you know, um >> that level of insight you don't have with humans today because it's impossible for a human to >> you know, like collect all of the insights they gather throughout the day, >> right?

31:32 And there's something interesting that happens when you start collecting those insights, which is not only do you have those insights, but now the full the the entire network of agents >> Yeah. >> learns about those insights automatically >> and it compounds. No. >> So, >> so yes, like to some degree, uh we're going to be that that in that insight layer, that level of layer of of um of intelligence, if you will. >> Yeah. >> Uh that's going to also enforce the next action. We we talk about this framework uh we call it a >> um the the action uh information um kind of loop.

32:09 >> Yeah. >> Uh the action that the agents are doing which is really where we where we come in first. We always start our our wedge if you will is >> starting with agents deploying work and and executing work. That's the uh the action piece, >> right? >> That action generates information. >> Yeah. that information you can actually convert it to an insight uh and use that insight to inform the next the next thing to do.

32:35 No, inform the next action. Example would be you know you're handling uh all of the inbound uh customer service uh for for you know one of these one of these uh say telos right >> and >> the information is well there's a lot of customers calling in about problem X. So the next action is deploying another set of agents that are going to follow up with the technicians and then like actually fix the problem at a at a bigger scale or like like notify the the the manufacturer that there's some issue with the with the router. I don't know.

33:10 So it's >> that orchestration that that um >> that you know connectivity layer if you will among your agents that is why we we think okay we're not just building like >> pawn solution for customer service. We're not just building a point solution for uh payment collections. And I think there's a space for for those point solutions. Yeah. >> Uh >> we just don't think that in the long run having that is going to benefit the enterprise versus >> being able to just deploy very seamlessly different agents that do different pieces of work.

33:41 >> Uh again, work is ultimately a combination of a process and some data, >> right? And what's that what that's going to create is for the customer to to kind of share all of those data points like all of that context in a way that benefits every agent, right? So again, it's very ambitious for sure. It's not going to be done like in the next uh like overnight. >> But um >> we do think that long term and we we we play a long-term game. No, like we we we see this as as an investment in the in the future of the company, >> right?

34:12 >> Uh but yeah, it's exciting. >> No, I think that makes a lot of sense. Cool, man. Well, we've got like five or six minutes left here. Um, I like to do just like some more rapid fire questions that are just kind of random. Um, first, >> you said the route. Sorry. Did you say the route? >> Oh, yeah. I said the route. Yeah. I said route for like 15 minutes out. I normally like to drive to Prescidio.

34:32 >> Nice. >> And then back to your office. >> Oh, that's cool, man. >> And then I don't touch the wheel. So, there you go. Yeah, >> it's good. This is probably the best investment. >> I'm going to get one of these for my parents in Spain. Hopefully, it works soon there. >> Yeah. I hope the I I wonder how the auto driving is is in Spain. >> Hopefully good because my my parents are getting all excited.

34:51 >> No, it's it's really good. >> No, no, no. Um yeah, rapid fire questions. First thing is that I like to ask is you know a lot of founders listen to this, right? Founding a company is hard. Um what is like what is like something that really bad happened that you're okay with disclosing and how did you guys recover from it? And for example, like LiveKit told us about like outages that happened and you know how they tackled that or just like any like crazy story that you're able to share.

35:23 >> Uh >> I had the the Telix guy was telling me that he got DOS attacked and then he had to like send Bitcoin to someone. It was it was crazy. But yeah. >> Oh man, let me let me think about it. I mean we've definitely had some >> some some fun like situations in the Oh yeah, actually I'll I'll tell this one. It's it's it's more of a stand it was more of um week-long process of like >> getting up to speed with the volumes we were getting. This is all this is 2024, >> right?

35:53 >> Um this is 24 24. We're basically fine-tuning >> uh Mistra and and and Llama. I think actually we ended up just fine tuning Lama Llama too or something. Um and I I literally bring my my wife uh who is a teacher. >> Yeah. >> And has vacation during summer. >> Uh to label like data sets like we we start labeling data sets and I I end up like >> becoming just like a like a data labeling shop for like a few for a few months and we're like five people at the time in the company. We're all like labeling this this one use case because obviously we were overfitting to to that one use case.

36:34 >> Um um and I think we just like labeled something wrong. >> Uhhuh. >> And I think the agent just like started like picked up on that and started going crazy on like a particular part of the conversation and then like this is >> this is maybe like I don't know a few thousand calls a day and suddenly like the customers are saying like >> dude like what what's going on? like it it just started like saying this stupid thing like I don't even know what it was. Uh >> um I think it it was it was actually uh some some like um something you know something the transcript actually was had picked up.

37:10 >> Uh actually this is the thing like the transcript from from from actually uh I told this in another in your podcast sorry in the interview you guys did uh with Dra. >> Oh yes yeah. the the transcript from Dig Graham that we were using at the time was changing the word, right? >> Was basically >> empty empty sound. It was saying something like, "Hi, this this is crystal." >> Yeah. This Yeah. >> Hi, this is Crystal. So then like we had somehow like picked it up in the in the labeling.

37:37 >> So So then like the the agent like was replying something super funny to like, "Hi, this is Crystal message." >> Yeah. They basically took us like another extra few days of like reabling the whole thing and like >> so uh I mean these sort of like hacky things they just have to do to like get >> it was our first customer our first real customer that we had really at the time like paying decent amount of money.

38:00 >> Uh we were literally labeling a data set >> for that one overfitted like >> llama tour agent. Uh so like you you do whatever you have to do now to like to get it done and obviously super like not scalable at all like to to maybe YC's feedback and recommendations like do things that don't scale >> uh obviously at some point you have to like build things that do scale >> uh and you I think you >> you have to be uh you know like very attentive and like be on top of things to like understand when you're going to do that transition >> but initially nothing really matters you know like if you don't have any customers yet Yeah.

38:37 >> Or maybe just have a few do whatever it takes. No, like >> go on side like figure it out. >> Maybe it sounds like not scalable, but eventually, >> you know, like you'll you'll get the signals that you can use to build something more scalable. So, >> no, I think that makes a lot of sense. >> It was not like, you know, maybe super like funny or like crazy, but >> No, I was, you know, it's >> absolutely >> every every week there's something.

39:00 >> It's a real chat. Makes sense. Yeah. What is what is your like take that you have that's maybe controversial that most people wouldn't agree with? >> I think we've had to to maybe um show this to to our to our network uh and to our to our you know partners in along this journey that it is possible to build a product that >> can cater to different industries. I think kind of going back to the other part like I think we're getting to a point where >> the verticalization is mattering m mattering mattering less and less >> in the product sense.

39:39 >> Yeah. >> Because it's so easy now to build you know like like a like a product like a UI per like >> um it's so easy to to do that. So >> where is the mode? >> The mode is only on the relationships that you build with that company on >> the word of mouth that you have in that industry. So >> yeah, >> yes, you have to kind of phase out your appro your your >> vertical like multi-vertical approach if you will, >> right?

40:04 >> Uh especially in in the enterprise. >> Yeah. Um but I think there's there's something there where >> you know things are changing so so much of the yeah from a product perspective like things are becoming possible >> that are enabling you know taking a bit of a of a stronger bet on on a certain approach right so I think that's like a little bit of something >> makes sense >> um that we've had to uh kind of convince ourselves really uh because initially it was hard like oh we're going to go like like target these these other this other industry.

40:37 >> Yeah, it's changing. It's like engineering, it's product, it's go to market, it's like marketing, like all of that stuff. Yeah. >> Yeah. Yeah. Yeah. >> Makes sense. >> Um, cool. >> Yeah. >> And final question here that I like to ask everyone is a little bit of a signature question. Okay. >> If you had a mascot, like an animal that could rep that would represent Happy Robot, what would that mascot be and why? >> I like the question. Um, I mean, we have a we have Nina, who who is uh Louis's dog, but that's like an easy answer.

41:10 >> No, no, it has to be something like something different. >> Um, >> I I think we're >> I think we're all like very like happy, high agency. Yeah. uh kind of overextim stimulated people like like we just >> maybe it almost like backfires sometimes like we we have like so much going on that that we're like so excited about stuff like we get very excited about stuff and >> um and we we try to like never lose focus but >> uh it's it's not easy sometimes and you just like see like a lot of noise right now lot of noise and and I think we I don't know maybe like a little um >> one of these like little marupil how do you call it the I'm thinking of a chihuahua right now. I don't know if that's what you're thinking of.

42:02 >> Maybe a chihuahua. >> I'm thinking of a chihuahua. Just like always >> super crazy. >> Uh no, I'm thinking of the little like um um monkey little monkey thing that in the desert >> or desert. >> The the one that lives in the desert. It's not a monkey. It's like a little thing that just like stands out and like looks around. >> Uh have haven't you seen one of those? >> Okay, I'm going to show you. Oh, you have a photo, right?

42:25 >> No, no, but like I can find it. Um, we can even like ask uh do we ask JGBT? >> A monkey that lives in the desert. I have never heard that one. That's a new one. >> Hey, what's the what's the name of the of this little monkey that lives in the desert and just or maybe not the desert, but like in the in the in the savannah. No, like >> just very excited. He just stands up and like looks around.

42:50 >> Oh, you're thinking of a mircat. They're not a mirat. They're actually small monguses. Mircats live in arid regions like the savas and deserts and they love standing upright to scout the horizon for danger or food. They're quite adorable. >> I don't know. I mean >> okay. >> Okay. >> I think that could be that could be it. But um >> also very >> very much of a of a um like you can you can sense the trust with with everyone in the company. So I don't know maybe maybe like a bit of a of a lion as well.

43:24 >> Okay. A mircat and a lion. >> Like a mix between a mircat and a liar. There you go. Very transfery folks. Like you can sense the >> you know like the team effort. Um >> I like that. That's those are very opposite, but they it sounds like they go well. Okay. >> It I mean it fits into our our maho culture. Like mahoo is a word in Spanish that means like >> nice and like wholesome, right?

43:44 >> Uh m a j o. >> Okay. >> Um >> it is [ __ ] crazy. Like don't get me like the the the company right now is it is it's crazy. It's nuts. is a lot of work. Uh, you know, like the bar is super high, but there's that sense of like >> we're all like having fun and maybe like like >> laughing a lot by >> building some agent that actually messes up uh w with with like some stuff and he's like, "Oh [ __ ] like that's crazy."

44:09 Well, the customer >> that sense of like, >> yeah, it's nice to work here. It's experimentation. >> It's experimentation. So like, yeah, maybe you're like all over the place, but then you you have that sense of like working together as a team, like >> going through the >> through the through the process together, which is fun. >> Um, >> that's awesome. >> Um, >> yeah. Cool. Well, thank you for joining. Of course. Yeah, it was great having you on.

44:32 >> You timed it so good, man. >> I did. It's perfect timing. >> It's perfectly timed. >> Now I've practiced enough where like I think I have a good sense for everything. >> That's so good. Yeah.