Hi everyone. My name is Patrick Akil and for today's guest I'm joined by Dennis the Rose, previously Head of AI over at ABN Amro and now Co founder of Stellar, where he works on AI for customer service phone calls. We discuss voice, technology, developer, productivity, output, startup environments and much, much more.
So enjoy. I want to get into what you're building now because I I know you have kind of this history built up and I don't know how much hands on you were, but definitely when you're now in a start up phase, I'm assuming you're a lot more hands on than you used to be. Yeah. What are you building?
So we're building Stellar, three of us at the moment, soon to be likely for Stellar answers, the phone with AI and we're focusing on people that have let's say at least 15 or 20 people in their contact centres all the way up to a few thousand. We have our first clients as well. So that's great. And am am IA lot more hands on yes, so. You have to be I. Started programming when I was how I know 12/13/14 or something. But I've never done that for
work per SE, but training. I'm a chemical engineer. I've done optimization of chemical factories. I was my own own consultancy business at the time, then went to McKinsey, did more factories, then with the McKinsey, pivoted into digital, then did the implementation work for the chat bots, then back into McKinsey, and then head of AI for AB and
MRO. So in all of those jobs, being able to program has been very useful because in the discussions with the engineers, like A, you can contribute to the discussion and B, sometimes there's sometimes just dynamic where there's some techno technical mumble jumble like now it's, it's very hard. And then you can say like, can you explain it to me? And then you can force the discretion a little bit more if you have some technical background. But now I'm actually coding.
So to answer the phone with AI, we're using the voice to voice models. That means streaming a lot of audio back and forth live. There's no text to speech in between, no speech to text. And then we have the guardrails running in parallel and so we picked go for the back end and I'm doing a lot of the codevelopment. I love Go Yeah. So my my 2 Co founders are more TypeScript. People oh OK, gotcha, but don't get around. Yeah, with with go, it's great.
And so especially like for the fact that we run a lot of things in parallel and I don't have to think about like, OK, so where's my threat located and the whole thing. So that works. So I'm, I'm writing more production level code at this point. Cloud is a huge help. Actually at this point it's Gemini. That's it's like it's kind of they they step over each other all the time.
So, but yeah. What's your, what's your, what's in your tool belt Then nowadays, because I'm very jealous, I think I feel like I'm missing out and I want to go more to building software in more early stage companies again, because I've also done product management for like the last year and a half and I always wanted to do product management. I'm very happy with the experience. I think I can leverage that in building software.
And I just want to again start doing that because I feel like if I'm not there, if I don't get the hands on experience now with the new tooling that is around, I'm going to be behind. Yeah. I mean, you have to get used to the tooling a little bit. Like what do I feel comfortable with moving into the AI or not? And how do I kind of prompt this thing? Well, I think that that's one thing that's been a bit of a learning curve in the beginning.
The other thing is that like some, some stuff works really, really well. It's like, hey, I have a bunch of functions here right in the tests and especially in something like go where you just have, you know, the under score test foul and it's like generate this foul for this foul and a poof. It's kind of 9095% correct.
I've had to force myself to be more careful on reviewing stuff and we do each other's PR reviews and as a result, you can, you can see that sometimes you pick something else like, well, this clearly is the AI writing this and slip through the cracks. Tool belt wise, we're using Cursor a lot. I've toys around with cloud code as well and I think they do slightly different things in a way like Cursor is a fixed fee amount thing.
So they have to be a little bit careful about how much tokens they burn in terms of staying profitable. And I'm sure that at some point they're going to raise prices on some things or there's going to be. Yeah, exactly. So that's on the curbside.
Cloud Code, on the other hand, doesn't do that because, you know, you just pay with your API key and they they've now done some stuff where it's integrated into usage of your Max subscription if you want to. But so Cloud is not Cloud code is not at at all optimized to reduce the amount of tokens. So for the same change in cursor, you could see like, oh, it's just taking, you know, the top 50 lines of the end of the
file. It's trying to find this function and then only takes the function into context. And Cloud Code will be like, I'm going to take these 10 files fully into context to make this. And, and at some point you can see if you do that, it runs out of space. Yeah, so smaller changes single files or something slightly across the code base. Cursor is great and if I something like hey, I want the thing to be slightly more autonomous and make a bigger proposal on what to change.
I tend to use Cloud Code from time to time as well. Interesting, what model specifically in in cursor? At this point, Gemini. Gemini, yeah. Gemini's king, I've heard. Yeah, yeah. Like, is it that significant compared to the others you've tried? I think it was just like SO37 was very good on the cloud side and then Gemini 2 1/2 Pro came out and it was slightly better and sort of switched over. It's trivial to switch slightly faster. The thinking mode tends to be more extensive as well.
Yeah, So it makes, it doesn't make a big difference. It makes a small difference, but you know, it's easy. It's a small step up. So you pick that. One that you say that because yeah, switching is easy and then you just use a different model, so you don't really think about.
It must be really hard as an LLM provider though, because even Google must be aware that at some point it will be a three eight, Yeah, yeah, the cloud version, if that's slightly better then voom, all that volume goes back again. It's a race continuously. It's like, yeah, you stop, someone overtakes you and everyone switches. Yeah. So that's very commoditized. So you think about it. That's crazy.
That's why, like for me, developer productivity and developer experience, those are going to be crucial. Not necessarily what model you're doing, but definitely how you do it and how you interact with it, how you achieve your goals. And I feel like people are experimenting with it. I want more people sharing knowledge on what works well for them and what hasn't worked well, because then we kind of established conventions.
Yeah, there's some. Really nice blog posts where people say like, hey, I've spent a few months with cursor and these are suggestions I make and like the obvious ones are to ask for a plan first before it executes. And you can see that by now they're optimizing the thinking modes to make a plan first before it executes. Of course, some of the best practices come back. And the other thing that's nice is you can define MDC files where you say like, hey, this always needs to be in context.
This needs to be in context whenever you're working on files that end in dot TS. And so you can basically say, I have my coding standards architecture of the application, etcetera, etcetera. And so it's easier for cursor in this case to tap into the right file or search for the right files because you're automatically providing that information. Interesting.
I feel like like I've talked to many people and a lot of camps are there people that love it, but also people that hate it, that say doesn't work, doesn't do what I want it to do. And I feel like this setup that you're talking about specifically giving the context or making a plan or making sure the right files are there, I feel like that is so huge and probably missing for the camp that's like it just doesn't work
for. Me, I think social people are sometimes looking for different things. So comes back to the beginning, like if I hire a junior engineer, I expect a certain amount of output, but I also expect to review the codes. Yeah, now suddenly I have an army of junior engineers that write code very quickly, but I still need to review it and sometimes say like this is not what I wanted. I actually wanted this the other
way. And the thing with an LLM said, unlike junior engineer, you know, it gets stuck as well. And somebody have said like, I'm going to reset this, I'm going to re prompt from the beginning because that's the better way to approach it. But so I think if if you come in with the expectation of this thing, it's going to write a lot of code. I still need to provide how I wanted written. Do I want some level of abstraction, abstraction or not? Do I want it, you know, split
across these files? So you have to think still yourself not to do it. And it's more the execution you automate. I think it works really well. But then I know people would say, like, I don't want to work with junior engineers. They need too much oversight. I prefer doing these things myself. So they have a different sense underneath and that camp of that, that group of people probably isn't super excited about the LLM for the same reason. Like they they're definitely not flawless.
They still need oversight. And if you if you come from the perspective like, well, it erodes my trust in that the string writes good code at all, then you might even I think it's hard to spend more time reviewing it than you would otherwise spend writing it yourself. But maybe you're very fast typer and thinker it.
Depends what you do like. I also feel like if you're if your code base is very extensive and you're trying to add on, I don't know if that's going to be better or worse, but if you start from scratch, you have to write a lot of code. New stuff goes really, really quickly. The interesting thing on existing code bases. So we're working with a framework called Encore and basically in the codes, both Azure can go, we can define like, hey, I have this database and I have each migration,
etcetera. And it sets up or whether it's AW or Google Cloud, but it sets up the environment based on what's in the code. So it's kind of an abstraction above, you know, writing the error form scripts basically just looks at the code, determines what kind of infra we need, and then it's been set up. Interesting.
And there was something in the command line tool that we wanted to adjust it specifically it's kept name for names when we would export from the TypeScript side a client for the Go sides. But Go has this thing with capitalised names being public and non capitalised being private. And so stuff that that would be public in the TypeScript side would suddenly be private on the Go side. Yeah, makes the clients really hard to use. So I'm like, OK, let's let's suggest this.
But the code base of of Encore is quite extensive. And so they were like, look, we're going to go and pick this up, but it doesn't have priority. I'm like, OK, well, you can make it PR if you want to. Yeah, great. Nice. And so that took like maybe an hour or two. But the fact is to be able like the guy who did that dunk case in the team, he's like that that would have taken me a week or otherwise. But now I can go open the whole code base and basically ask the AI like, OK, so this is the
change I want to make. Where do I even start? Yeah, like, OK, let me check. This is the structure of the application. It's likely going to be here. Let me show it through these files, OK. This is structure like this. No, Then I need to be here. Yeah. This is the function where it happens. OK, make the change makes the change change the show, show, but still you've you've gotten to the right point. Yeah, you can edit basically your approach and there's some back and forth on the PR and
then goes merged. I love that. I love that. I hope people use this also then to contribute to open source, right? Because if you need something, if it's crucial to the core of what you're building, it's a dependency. And it's super nice that it's open source, because then you can do this. Yeah, and it's like usually the burden to get into a new code base is quite high. And for making like a single change that you really want,
suddenly it's very low. If you want to be a core contributor on something, you still need to really understand the code base, but in this case you can ask AI to get you where you want to be. I wonder if what language was this in Was also Go? Yeah, I wonder if that make is going to make a difference. Like if languages are going to define then also what works well with reasoning with an AII? Think maybe it's less languages. I think it's more time bound.
So if a language like Go has existed for like what, 1214 years? Yeah. And there's enough training state and the same for type grid and a bunch of other things. If I now start a new language, even if it's a great language, there's not that much example code. You can't catch up. And so no. And the same goes for frameworks, right? What if I invent the next Ruby on Rails? There was no way that the LLMS are going to be able to pick that up easily.
And as a result, anyone working with AI is going to be like, yeah, I can't work well with this framework because it's too different. Yeah, it's really funny because I had this conversation with a person that loves Rust. I was like, Rust might be one of the last like established programming languages then, right? Because of this, if you have a new language, can you play catch up with all the tooling that is around and with people feeling productive and effective? Like how can you?
Yeah, it's really hard. And I could imagine something which said like, well, for this language, I'm going to define your cursor rules files that, you know, spec out the language and how to use it, the touch pad. But I wonder if you get to the same performance on that than basically being able to learn from, you know, years and years of GitHub. Repositories.
Yeah, I went to Japan and then I, I know a little bit of Japanese and then I had this thought because we prompt and we do everything in English, but is English actually the right language? Like is it the best language if we say the word run? And this is what because I, I chatted about this and Jajibti told me if you have run in many contexts, it has like 60 meanings. Whereas a different language like Japanese, you are more explicit in the words that you use.
And for me, that's also where the thought pattern kind of went into programming languages. Because if we look at Go like you have to be very verbose. You can't do many magical things. And in the end, there are a few ways to get the same thing done. You still have flavours and you still have conventions. But then that's that.
If I then look at Java, you can do many things in different ways and you can be very opinionated or you have to have strong conventions to then actually have consistency. Same for Rust. Same for Rust, like so then I feel like a language might make a difference, but in the end, yeah, it might not. And then these are just the languages we work with. I think Go is a very simple language, like very few keywords, yeah. And so probably works relatively well for the LLM scope.
The other thing I see is that on the TypeScript side, we do a lot of the kind of the correct operations. And the Go thing is slightly more complex because it's a lot of parallel processing of audio and so basic correct operations. There's so much training data on that. Even if it's not, I mean, it's also in TypeScript, even if it would be in different languages, I think it it works well. The more complex code, it tends to restore a little bit more.
Yeah, gotcha. Do you, do you still then try like because you're doing audio to audio processing basically that's like, I wouldn't say it's not established, but it's definitely something I haven't done before. So it's in the newer side of things. Yeah, the other thing is that it turns out like even very trivial stuff you'd say. Like I have raw audio recordings in some formats, PCM 16. I want to turn this into MP3 or something. There's very little available to do that. Yeah, yeah.
Like there's a few of the codecs but they're all C related so you have to tap into C, or course, yeah. It's not like image conversion. No, no images. We use everywhere. It's interesting to kind of even see like, hey, how many libraries are there? Like how many people do actually do kind of significant audio processing and go, it's fairly limited, I think, but there's not that many people that do a lot of significant audio processing, especially in the open source space.
Yeah, on that front, you're kind of trailblazing. But then do you use AI to help figure things out, or is that where you kind of have to do more of the manual plumbing yourself? And so I think the set up of like, hey, how do I want to route all this? How do I set up the go routines? How do I run my guard wheels in parallel? That's something you map out and kind of here white boards or paper drawing.
And then it's more of like, OK, so I have this vision now I would like to implement, you know, this speech or this speech or this encoder or something else. And then sometimes cursor goes overboard, which writes a complete, you know, VAF encoding thing. That's like OK, there must be a library. That's like, yeah, where did you get this from in the 1st place? Yeah, interesting. I was talking to a friend of mine, Zoe Langdon, and he's also doing a start up.
His is more OK from a learning perspective. I want to learn about cloud topics and his app is going to help with regards to centralizing and providing high quality content and kind of a learning journey. And he says, well, if I'm, if I'm vibe coding, usually I need to nail whatever I'm trying to do in one or two prompts. Otherwise, I'm going to start from scratch. This is what I meant with the IT tend to get stuck.
Like at some point you have so much of the kind of wrong direction in the context that it's you can prompt say like, no, I don't want this, I want this, but it's kind of it's still overweight for everything that happens before and you're better off. I mean, maybe not 2, but after three or four prompts, at some point I have the same feeling. It's like maybe I'm just going to go and and reset this and
spec out from the beginning. Better what I actually want and then it can take the right route. So you have the same feeling like a few prompts and otherwise you just. Yeah. Once you're in the weeds, you need to get out first. Yeah. This like I haven't done this much and especially not on a production application, but this feels to me like a video game and you're trying to get to a next checkpoint. That's also what I said to him.
It's like it's very much checkpoint based, like you have a good safe state, you better commit and then you go again. Yeah, yeah. But it's kind of I, I think it, yeah, I think the analogy works in the sense that like with the junior programmer, you, you want to get to a certain point, you scope out the path to, to get there. Yeah. Sometimes the eye comes back with like, hey, you've forgotten to think about these and these and these things like, Oh yeah, great.
That's something that otherwise I would have run into. And then you then you move. And I think as the AI matures, these checkpoints will be further and further apart. But I think you still need to think about the route that you want to take and how do you slice it up in something that the AI can follow. Yeah, I want to talk more about the product side and more of the business side of things because from my perspective, I don't know what it was.
I wanted to return something or or ask a store when they were open and I was just on the phone for hours. So it's like this should be a solved problem, and I feel like you're getting close to that when you're. Talking about customer service, So you probably want to return something because people tend to call when it matters or when
they are not that easily savvy. But there's a fairly large group that picks up the phone when something is important to them and a return is more likely to be important than the opening hours. So, yeah, look at at ABN, we introduced voice for the credit card company and it's, so it's
in trials. It basically comes up to when you call and there's a waiting line and AI can at least, you know, ask you what you're calling about, identify you and then do it or I'll hand over to A to a human agent. And of course, over time you could do more and more with that. Now last year, May, so roughly a year ago actually think about it opening showed didn't release at that point the real time audio functionality, I was like, OK,
so this is great. It can do sarcasm, it can sing, it can do a bunch of other things. And a lot of that is based on the audio to audio approach that that they've taken. So I think the API became available in October, September, October last year. Yeah. And I think that was a moment that I thought like, hey, finally we're going to make this switch. So up until that point, voice bolts where exactly? What you feel when I say the word voice bolts, it's like it's very robotic.
And you know, it's a bot. Yeah, you know, it's a bot. Yeah. And it's also it's it's like super tiny scopes. Like what is your first name? It's Dennis the Dose. Like, so what's your last name? Dennis the Dose. Next question. Yeah, yeah. OK, that's the next question. Let's continue. And so it's, it's like it's very static, but also very impersonal. And suddenly you could feel like, hey, this can now pivot into something where it's a great experience.
And so as an example, I'll get the example in a bit. So that that's one for me. It clicked like, hey, there was something here, it could be much bigger. And that's when roughly December, I decided to leave ABN to go and build this company together with the, with the other two founders. Now as an example.
And last week we were in a discussion with a board member and his management team from one of the larger banks in the Nordics. And we explained that we've done tests where we asked people like, hey, look, you've now been helped by the bot. If you had to wait for 5 minutes and then would speak to a human directly, would you prefer this experience with the bot or would you have preferred the, the weight and then the, the human connection?
98% of people say that they prefer to speak to to this bot particularly. And so good stuff. Yeah. So one of his empty members on the table, I think the chief data officer basically says like, well, I would be part of the 2%. I would never want to speak to a bot. And so we did the demo. And the demo is basically somebody that's calling about, you know, I, I don't recognise his credit card transaction. And then the bot is not just saying, oh, I will, you know, return it to you.
It's like, it looks like this other transaction. Could you maybe recognise it like this? Could you, could you check this or that? And so we solved this. And then at the end the same person looks at me and says like oh now I get it. I would prefer this bottle for human any day and. Yes, that's exactly what you want to hear. That's what we want to hear. Yeah. But I think it also shows that kind of it's pivoted.
It's pivoted from the robot where you feel that this is a cost reduction measure to, hey, it's actually very natural. And suddenly like instead of being open only five days a week from 9:00 to 6:00, we're open 24/7. We can help you with all the basic questions. And so one of the things that we're getting a lot of response from at the moment is healthcare. Yeah. Whether it's dentist or dental practice or like GPS or the house charge protect General practitioner. Yeah, yeah, I know.
But it's the the central thing where you go if it's out of hours. And so they have people on the phone and it's not a great experience. It's very hard to plan for this. And it's like, OK, so we would like this, but the the original way it would work would could never succeed in such an environment. We think this can. And so we're doing a bunch of FOOF concepts there as well. Interesting. Well, yeah. So I think this is the big
thing. Like it suddenly doesn't need to be this robotic cost saving approach anymore. Can be really a great experience. What are you focusing then on? Because like for me, if I call indeed my General practitioner, I'm going to be in waiting line. Basically I call as soon as they open and still I'm #10 in the waiting line. That's one angle. But then you also have the more massive, more e-commerce angle. Like are you tiring to Tate or both? Or are you actually saying this one?
First, we're going to nail this one. Then we go broad or. We're very much focusing on the larger parties. Yeah, let's say at least 20 people in a call center or so. There is some scale in healthcare though, like where you have, if you have a bunch of dental, dental clinics together, you could say like, hey, I can make one AI that can answer the phone for let's say 100 of those practices. And they all have like, you
know, 0.5 FT on the phone. So if you centralized that in that way, it could still reach the scale that it works for us. Yeah. But we do try to, we give a very personalized experience and that means investing time in the bot and we tend to integrate into a number of systems, which again
is an IT investment. So this, this, I think there's a segment below this where you kind of put in your credit card and it's kind of works, but it only integrates with Google Calendar. That's not what we're aiming for. Got you. But then you haven't specified like, OK, we're doing this domain and not these domains. You're looking at how many people are actually then doing this in the. 1st place gotcha. And so a lot of the product is how do we handle the
conversations well, safely. And so we, we do still have steps in the conversation, but they tend to be much bigger. So rather than the step that says what is your first name, the step is more like introduction, identification of topic. And then there is a step around identification of person. And then there's Step 4, you know, in general questions, Step 4, moving an appointment, making appointment, these type of things.
Gotcha. And there's 0 user interface like from from that one FT that's still there. You don't need any user interface, do you? So, yeah, the client interacts with it on the phone, Yeah. But there is an interface where you can kind of you edit the conversations, you can monitor, you can listen into conversations, you can take over a conversation if you think like this is not going well. Yeah, there's a lot of guard
railing around it as well. And so you specify the guardrails and the tests of these guardrails so you can automatically test and retest all your conversations. Integrations is a thing like how do I make API calls? How do we do the authentication on API calls? How do I do handovers? Yeah, gotcha.
So yeah, I mean, there's sort of the AI and of course, that's a very complex piece of thing, but turns out that to do the set scale for larger companies, yeah, everything around it is quite substantial as well, of course. How do you store and retain and process data all of these things? Yeah, it still has to be a company. Yeah. It cannot just be the piece of software. You have to build a company. Yeah, Yeah, that's interesting.
Now then you mentioned you started let's say half a year ago or at least something came out half a year ago and then you started 445 months ago. How far are you now? Because also you've been building with three people. You're going to hire a fourth person. You I think consciously haven't ramped up in people yet.
Is is there no need to? Well, if we had a bunch of fishy money, but we've deliberately chosen not to, I think we would be hiring more in the implementation side than on the building side, OK. And implementation you mean by doing client projects? And yeah, so sales, we do a lot of ourselves, yes, it's also good to get a sense of what do people actually want to buy, what are the problems they're having. But once we do a project, of course you have somebody there that says, OK, I need this
project. I understand the analytics of the type of goals we want to automate, the systems we need to integrate with. And so that type of engineering, the kind of more integration engineering is slightly different than the kind of building the product. Gotcha. But yeah, we're very productive on building the product. I would say especially in the first two months, like things went so fast. And I think we discussed this a little bit in the in the pre
call. But like if you look at the code base at some point and you can do this analysis based on the Kokomo estimates of how much, how many people would you need, how much time at which cost to build something. And then it has a dimension of like, well, if you do this very waterfall, it would be like this. And if you do it a little bit more organic, it would be like
this. And if you do that type of analysis on our code base, you get to something like, oh, you know, this would take you a year and a half and like 7 people. And we did it in four months and three people sort of. And part of that is the benefit of being early. And I used to be very critical of those estimates, but I've done this in ABN in the context of not using these type of tools in a little bit more of a corporate environment.
And then they tend to be fairly on the dot where it's like, oh, this would have taken, you know, 12 months and six people. And it's like, OK, fine, we took 18 months and four people, but it multiplies to the same same amount of months. And so, yeah, I think we're going very fast. How's it feel? It must feel good, no it. Feels very free, yeah. A free. That's the word you choose. Yeah, Yeah.
That's awesome. Look at if you're in a regulated environment, there is whole departments to manage risk and the best way to manage risk is to sometimes not do things of course so. You don't have to. Yeah. So I mean that my previous role was much more of a diplomat, I would say, and you have to get people to align around like this is actually a good thing to do. Yeah. I mean, I think head of AI at at a bank, specifically ABNM role, like you have to be a diplomat
to a certain degree. You can't just be a person that is continuously building and and has a high risk appetite. It's because a bank can't afford that. True. But at the same time, I think if you're in that type of role or innovation more broadly in this type of environment, you're kind of expected to push the boundary a little bit. The question is like where to
push, how far to push? Like if if you're, you know, a three time founder, very successful entrepreneur that's always done this from scratch and then you go and try and renovation and banks going to be very hard because there's some stuff that that just doesn't work in a corporate. But I think it's an interesting role. So you you're continuously fighting the system and it's somewhat exhausting, but it's also somewhat kind of, you know, comes to the territory of that job.
Yeah. You mentioned earlier no VC money. No, no capital injection. Is it no capital injection yet or do you think this is going to be a trend that people uphold because they can be more productive with a smaller team, so they don't necessarily need to hire or kind of blast for the moon, They can do it kind of organically. So I think what you can do today is you can get much further with the smaller founding team and so you can delay the decision on whether it makes sense to to get
funding. What I see now if I look forward, if we start converting a lot of our client projects, you have a moment where your growth kind of outpaces your cash flow And, and at that point you kind of, you know, are you going to a bank loan, but the banks are not going to be super easy to you know, fund the company. It's only been in existence for a few months. Yeah, risk again. Yeah. Are you going to kind of the VC appetite?
But I think going into VC money comes with an expectation that you're going to be, you know, one of the companies that's going to give a 10X or 100X return. Yeah, you have to be a multiplayer, yeah. And so that's a decision as well. Like the moment you go there, it's not just money for this, this phase of growth, it's money with an expectation of, you know, hyper growth afterwards. And so for us, like we wanted to prove that we could sell this, which we have, that the product
would work. And I think in a few months we're going to have much more data for ourselves as well to make a decision on on how to proceed after that. Gotcha. Yeah. What's this what you envisioned, let's say 4-5 months ago when you had this idea or when new technology came out? Or did you have to pivot and is this where you ended up? Well, it's not that black and white. I think any entrepreneurial adventure starts with the assumption of there being demands for what you thought.
About That's a big one, but you have to have. That, but that that's kind of the, the, I think that's that's where everybody comes from in an entrepreneurial space, like, hey, I think people want this, whether it's AI consultancy, whether it's a product or something else. And so you start your discussions with clients, I'm like, hey, we have this thing, we have this vision, this is what we're building. Here's a demo.
What do you think? And I think in our case, we've been fairly lucky in terms of, yes, there's demand for what we had in mind. But I think if you and this is the natural black and white part. If you look underneath that, we have really made changes in our approach of like, OK, so it turns out that this is much more important than that. One of the things we pushed forward initially was saying integration to phone calling systems like Avaya or Genesis or said like, look, we're we're
it's a problem. It's it's a technical thing to solve, but it's not that challenging. And so we'll pick that up when we get there. Yeah. What I find in client discussions is this is one of the things that's most open mind. Do you integrate with the system? And so it's something that we're pulling forward. It's not a pivot, but it's like you're continuously adjusting your road map a little bit to what you think is more or less important.
Yeah, gotcha. Yeah. I, I also think they're different than a pivot like learnings along the way or what is going to be more effective for the use cases is going to always be there. But I feel like this vision of, OK, we want to, I don't know, improve customer delight by leveraging AI for XY and Z like that. That vision's still strong.
And I feel like it's something that even without like the latest technology, it is already what companies were trying, right, Automated answering systems, OK, dial XY and Z to get to this department faster. Like we were in this accelerating process, but not to the degree that we can do now,
which is really cool. And I think there are a lot of companies that want of a certain level of great experiences with our clients and, and have a good interaction because it fits with who or what they want to be as a brand. And so for them, it was very hard to put something in there that says press 1 to do this, press 2 to do that or to put like, hey, you're going to be like the 10th person waiting.
And so it's for for that group, it's been very hard to do something in the past because the technology just did not meet, you know, any of the basic requirements or the hygiene factors of how they want to have the phone based interactions. And I think that that's now suddenly coming into scope. And then there was a group that was already doing this, but they were coming into it from like, you know, this is the the one way we have to manage, you know, call inflows, etcetera.
How do we now make that better? And I think then also something like Stellar comes into scope. Yeah, nice. What I have seen and that is more with digital marketing is that user demographics or an audience, the more data that you have, the more you can tailor the content towards that
audience specifically. And for me, it's interesting because if I look at kind of a phone conversation, the person on the other side almost has no idea who I am right from the initial, OK, I'm calling when they know my order, they get my information. So then also if I'm looking at AI, you would be able to tailor kind of tone of voice and conversation towards that data, but it's like a chicken and egg problem because you don't have it yet.
You have to have minimum data to get more data from the systems that you have. I, I think what works well though with the audio to audio setup is where I said like look, try and match to some degree the emotional tone or the type of interaction you get from the other side. So if somebody's a very fast speaker, you may speak faster.
If somebody's, you know, seems to be a slower speaker or seems to be emotional a little bit on that, I'm like, don't try to be overly cheery if the other side is not overly cheery. So I think on that level to be able to match kind of that emotional response, I think is is a thing that you can already do quite well. And then the question is like, do you really want to, based on the profile of the person calling, change how you interact? And there's a creepiness aspect to it.
But the other thing was like if I'm a, I'm the same brand that wants to be fairly cheery and outgoing, etc. Do I want to use the formal you or the informal you or do I basically? And do I change that depending on the person on the other side to say like, no, our brand image is that we are young and cheery and therefore it's gonna be the informal form all the time. So it depends a little bit, I
think on how you approach that. I, I didn't think of kind of using the caller's voice and their, their tone of voice or how fast they speak like everything with regards to their style of communication in the direct response. And it it does that accurately like it can, for example, if I'm furious, it can kind of tone it down and take a a more careful
approach. So if you're very cheery, it might be very cheery, but like we, we ask to try and match emotion, but don't be, don't be negative, don't do defensive. You're almost prompted out. Sorry. If you're very cheery, it'll be relatively cheery back. And if you're very negative, it won't try and be cheery. It also won't be negative. It will be like, oh, I'm very
sorry about that. That's the and then initially we had a bunch of guard wheels trigger triggering and like, oh, you're saying sorry that that could be an admission of guilt and that might be negative to the companies. Like OK, apologize for something is OK. It's OK. It's not that hard. Yeah, Yeah. You know, that's funny. Like we're now gonna. I mean, I'm assuming you're doing this in the English language, but eventually if this, that's a big. Assumption, but that's not the
thing. Oh, really? You're also doing multiple languages? Yeah, OK. So we're doing most of it in Dutch at the moment, OK. But we can do German and Danish and a bunch of other things as well, yeah. Interesting. So you're not bound by a language? No, that's very cool. Yeah. I mean, it's going to be hard to do something very niche. I'm not sure how we would do in Swiss German or something, but all of the major languages like French and German, it works really well.
Dutch works really well as well. That that's for me is kind of strange because I wondered then like in what on what data is it trained specifically? Like the linguistics and specifics of a language? English is there in abundance because we have incredible content, but like, is it also then trained on make a share? You have Dutch television like it must have been trained on a lot of that.
Yeah. The other thing is, I think in the end, the model abstracts and so if I really dumb, but then I suppose the embedding of the same sentence in English and Dutch are relatively close to each other. And so you can have something pronounced in Dutch which only has, you know, English based roofs, what it was trained on and it will still likely work. So you I mean, it's not that they have separate models for Dutch and English. A lot of that knowledge blends together and still works.
There's some, like one of the things that's interesting where you can see it go wrong is that in English would say 82 and in Dutch would say yeah, 2 and 80. 2 and 80 and that's just like the only one that does that the. German shoes as. Well as well, yeah, I saw like, I don't know if it was Danish, but they have something incredibly difficult where that's like and the French. As well, of course, if it's like catrifandis, it's like 4 * 20 + 10 exactly or or also known as 90. But yeah.
And I think the Danish have like something that's like base 16 or. Something that's the worst. Yeah, it was worth a. Programmer's perspective base 16 makes perfect sense, but no, but so you can see it makes mistakes there pronunciation of numbers to where the Dutch Schwalen is switched around first, the English 1.
You can see the English training states are kind of pushing, but I, I suspect this, this is one of the things that they're working on really hard because it's kind of it's common. You can see that almost every in a lot of the Germanic based languages. It's really funny because that is also if humans are, if people are learning a new language, that's also what trips them up. It's like, oh, you why you switch this around like that doesn't make any sense.
Yeah, that's pretty funny. You mentioned you're going to do a first hire. You're growing from three to four. Yeah. Is that the first hire or? Technically speaking, this is an intern as a working student gotcha thing. But yeah, yeah. And that links all the way back to the beginning where like if you have young eager people with the right amount of drive, that will work. Good stuff. Because I was wondering like is this a decision or but this kind of grew organically you started
with? An internship, yeah, we're also talking to some people or to like, hey, if we converge a set number of projects, yeah, these would be great fits from our network and they are more experienced to kind of to join the company as well. Gotcha. So it is all network based currently. Yeah. Like in that small of a stage, like I feel like we talked about hiring, it's very much different from like the US culture versus European culture and that's going to impact also your start up.
Yeah. And the yeah. And I think the ability that we have with AI to kind of stay smaller for a longer amount of time, but client implementation projects need FaceTime and and boots on the ground. And so that's going to be something that we eventually need to hire for it. And we're looking at partners as well. So we could work with different consultancies which are like, hey, like the first few we have
to do ourselves to learn. But some particular said like, OK, so this, this is how an implementation project works. This is the handbook. We can start 5050 with a partner and at some point you could staff, you know, 8020 or you could fully have the implementation. We run by partners that are experienced with the product. Got you. Yeah, as last kind of thought pattern, because we we talked about where you started, where you are now, it feels like things are accelerating.
Where do you think this is going to be like end of this year or will you be at that stage with very much like hands on work on a product where you're confident and you're going to do more implementation projects or what's going to be that phase you think? I think if we do well though, I mean, no big company is going to say, hey, here's my credit card. It's not going to be a credit card anyhow, but here's a here's a contract. Take my money.
And we're starting the 1st of June, let's have all our phone conversations answered by AI. There's a few people that we talked to that are, I think a bit more aggressive. And I don't want to approach it. But by and large, they're going to say like, OK, you know what, go and build the first use case with us. And we're going to go and test that with some people and maybe a client panel.
We can do this in a few weeks. Then afterwards, state pilots for, OK, let's get actual client calls, but maybe just Route 1% or 10% of the calls. And if then we're so confident that's that's actually scaled this up and even then it's not from 10 to 100% all at once, but you know, slowly increasing the volumes. So I think a lot of it is around people getting comfortable if you don't plan that on the road map towards the end for the year.
We're now starting some of these proof of concepts on pilots. I think if we do well, we're going to go and have some production implementations in Q3Q4. And I've been to the end of the year, we'll have, if all goes well, a number of these implementation projects running next to each other, fed by a number of concepts and pilots with new clients. At that point, we're going to be slightly bigger company, but I think that would be the ideal
outcome. And then we're focusing not just on the Netherlands, but more on the European market. The US market is special at the moment with the whole administration there, but also the we're trying to distinguish a little bit on not the English languages. Gotcha. That's right. There's a lot more players in the US, but they're focusing only on English. Gotcha. I love these intake forms. It's like, do you serve your customers in English? Like, no, pop us up.
A new question. Could you serve them? No, we'll put you on the wait list. Oh wow, yeah, interesting. I I was wondering because another thought just popped in my head. I know a lot of people in the start up ecosystem and some people when it it's about a new technology, they go stealth mode. They don't come on a podcast, they don't talk about what they're building because they have this fear of other people outpacing them. Is that a thought you have as well?
Because you are very open with regards to this is what we're doing, this is what we're leveraging. Maybe not the intricacies of how exactly, but I love that. Sharing this. So I mean, some people might have good reasons to be in stealth, but I I think like the barriers to compete with this issue that needs to get to the table with, you know, board to board managed one type of positions in large corporates in the Netherlands. And that's not something that you can steal.
Like if you're going to like, we're going to go and talk to the person, you still need to get the meeting. So I believe a lot more is in execution than in kind of hiding what you're doing. Intellectual property, yeah. And of course, like what we do in the code and how certain things are structured, etcetera, there, some might be in there. But I think if somebody says like, oh, this is great, Adi, I'm going to go and start a voice.
AI started trying to compete with Dennis here, then I'm not necessarily too worried because there's a lot more to it than just building the product. Gotcha. I love that man. I think this was a really great conversation. Thank you so much for coming on as well. Is there anything you still want to share before you round off? No, not specifically. Good stuff then, thanks again. I'm going to round it off here. Thank you for listening. The best way to support the show
is to leave a comment. Let us know what you think of this episode. Like the episode? If you liked it, subscribe if you want more and we'll see you on the next one.