Luke: [00:00:00] You’re listening to a new episode of The Brave Technologist, and this one features Riccardo Arnaldi, AI Product Manager at Pandora, with 6+ years building and iterating on digital products through data, experimentation, and AI. He’s currently owning end-to-end development of strategy of LLM-based shopping and customer support agents at Pandora.

In this episode, we discuss balancing AI agents in a high-profile brand, having a seamless experience between service and commerce agents, along with interesting lessons learned from having agentic AI in production. And now for this week’s episode of The Brave Technologist.

Riccardo, welcome to The Brave Technologist.

Yeah. How are you?

Riccardo: Good. Good,

Luke: Good. We’re here- It’s a good conference … at the AI Summit in London, did you do a keynote today?

Riccardo: Yeah I’ve got a keynote at 2:30-

Luke: Oh, yeah … on the New York stage. Well, what, what are you gonna cover in the keynote?

Riccardo: Yeah, so I’m gonna give, like, a overview of how we went from, like, having a product that it was really, like, only in staging and a vision of the company, so having a shopping agent, which is something that is not super common, into productionizing it, [00:01:00] and now having it, live in a couple of countries very soon.

Like, it’s live in Australia, and I’m gonna be live in, uh, UK at the end of the month.

Luke: Awesome, awesome.

Riccardo: So it’s, like, getting some traction on that, and how we moved from something that it was, yeah, not really great- … to something that we can be proud of having in production. Yeah.

Luke: Yeah, I mean, like, it’s, I’ve been looking forward to having this conversation, ‘cause, you know, people, you’re seeing bits about agentic commerce and, and shopping and all that, and obviously, like, is such a huge part of everybody’s lives too.

I mean, when you think about measuring success for a shopping agent, what metrics are you leaning in on for that?

Riccardo: Yeah. So if you ask me as a PM- Yeah … you’ll get a different answer if you ask to, like, the business side.

Luke: Let me ask you- Yeah … as a PM first.

Riccardo: Yeah, exactly. So as a PM, I really look into the main features of the product.

In this case, it’s a shopping agent, so it needs to, you know, be able to reply to FAQ questions about the products, the sales, the promotion, and mostly what it does is that it has a very strong RAG, right- Yeah … so retrieval. So it needs [00:02:00] to be very good at bringing products to the customers, asking, of course, using natural language, right?

So the measuring of that, we usually use helpfulness as a score.

Luke: Okay.

Riccardo: So basically we look the multi-turn conversation and we see, okay, is the multi-turn helpful in terms of, like, is the agent answering in a way that the customer finds helpful?

Luke: Mm-hmm.

Riccardo: And you can determine that by using, LLM as a judge with some instructions.

Oh, okay. So it will tell you, okay, this conversation overall was good or bad. We have a score, from zero to one, one being, like, super good, very satisfactory, and zero being very poor, and then we score that in different fields.

Luke: That’s really interesting- Yeah … ‘cause you hear about judging for, like, benchmarking and things like that, but I haven’t heard about it for, like, a helpfulness rating. Yeah … there a bunch of optimization you have to kinda do to get the judging- Yeah

dialed in? Yeah. I would imagine, right?

Riccardo: you got to have a human in the loop-

Luke: Yeah,

Riccardo: yeah … that reads, the conversation, kind of evaluate it themselves-

Luke: Yeah …

Riccardo: and then try to tune the instruction of the LLM as a [00:03:00] judge-

Luke: Yeah …

Riccardo: to obtain kind of, like, the same score.

Luke: Yeah,

Riccardo: okay. Right? Cool. So it’s a process.

you’re not gonna get it right at the beginning.

Luke: And is the agent, like, primarily working from, like, the Pandora domain, or is it the agent kind of working across the broader web?

Riccardo: No, no, no. So our agent is deployed on the storefront- Okay, cool … on the UK.appaganda.net.

Luke: Nice, nice.

Riccardo: So it’s there.

Luke: Awesome, awesome. Yeah. how do you kind of run AB tests and iterate on an agent when the output is non-deterministic in people’s conversations? At least people think that their conversations are very unique. How unique are people’s conversations, right? Like…

Riccardo: Yeah. So they are pretty diverse.

Luke: Yeah, yeah, I

Riccardo: would imagine. Also, like, in the way, maybe you ask the same thing, but you ask it with different words.

Luke: Mm-hmm.

Riccardo: And we know that LLMs are very, like, semantic, aware, so you will get different answers- Yeah … as you say. So there is a few ways that we are using to govern that a bit, like how the agent is supposed to answer.

Also because our business stakeholders, they really want something a bit locked in.

Luke: Yeah,

Riccardo: yeah. So we cannot have, like, LLM going rogue and giving, like, very [00:04:00] different, uh, answers.

Luke: Yeah.

Riccardo: So recently Agentforce, we have Agentforce as our provider. Oh,

Luke: okay,

Riccardo: cool. Right?

Luke: Nice.

Riccardo: So Agentforce released something called Agent Script-

Luke: Uh-huh

Riccardo: which is bringing a layer of NLP-

Luke: Oh, okay …

Riccardo: determinism to- Yeah … the actions. Okay, cool. So one of the bigger tips that I will also give today in the talk is that you can add this determinism layer- Uh-huh … on top of LLM. Mm-hmm. For example, to do some routing, sometimes, I don’t know, if in the sentence the customer writes, “What’s the warranty?”

something, right? Mm.

Luke: Sure, sure.

Riccardo: The word “warranty” will be intercepted by the NLP-

Luke: Mm …

Riccardo: processing, and then the routing is gonna go through FAQ. Oh, okay. So the agent LLM is not gonna decide what actions and what tool to use. The tool routing is already done.

Luke: Interesting.

Riccardo: So then, you know, we’re sure that the output is gonna be FAQ answer.

Luke: That’s cool.

Riccardo: Right? So-

Luke: That- that’s really interesting …

Riccardo: it gives some determinism to that. Yeah,

Luke: yeah, yeah. Yeah. That’s awesome.

The Brave Technologist is brought to you by the Brave Search API. Access billions of indexed web [00:05:00] results from a simple API call with the Brave Search API. Join the leading names in AI and tech using the Brave Search API to power agenic search, keep LLMs current with real-time data, train foundational models, and bring the best of the web directly to the leading edge.

Get started today at brave.com/api.

what’s kind of the most underestimated challenge for companies trying to build customer-facing agents that actually work in production?

Riccardo: Yeah. I’m gonna give you the honest, a bit boring-

Luke: I

Riccardo: love it

uh, answer, and it’s QA.

Luke: Yeah, yeah.

Riccardo: So, yeah. Especially in bigger companies like mine, we have a army of QA people- Yeah … often external, and they think they’re just gonna apply the same, you know, principles that they had. But it’s an agent and it’s not the same.

Luke: Right, right,

Riccardo: right. Right? And you cannot automate it with a playwright script.

Luke: Right,

Riccardo: right. You need to put some agent LLM-

Luke: Yeah …

Riccardo: to test the a- agent. Okay.

Luke: Yeah, that’s what I was gonna ask- Right … is, like, the QA, is it, like, is it all human QA? Is it kind of a mix of- of human and [00:06:00] automation, or?

Riccardo: It is a mix, and at the beginning it was mostly human.

Luke: Yeah.

Riccardo: Then we asked our partners to help us-

Luke: Yeah

Riccardo: automate. They didn’t really know how to do it Really? Really? Because it was, like, one year ago- Yeah … and it was very new. Yeah. And Salesforce was like, “Fuck, we need to- … to move from Salesforce to Agentforce.” Right. But, you know, it was a transition. Right. I think now they’re getting there.

Luke: Yeah.

Riccardo: But luckily enough, we have very good engineers also in QA.

Luke: Yeah.

Riccardo: So we managed to build our own pipeline- Mm … to run the QA.

Luke: Yeah.

Riccardo: And now it’s mostly automated uh, using Devin. So you have kind of like a testing pipeline, single turn and multi turn, and that’s very important because otherwise the developers will build a lot of features, but then QA cannot sign off.

Yeah. And of course, a big company like us, we have millions of customers every day, you cannot just ship something that it’s somewhat QA.

Luke: Right. Right. Right? Especially on, like, high-end product, right? Yeah. Exactly. Like that you guys do. There’s the expectations from the customers around quality. We

Riccardo: have a brand reputation to protect as

Luke: well.

Exactly. Exactly. Yeah. No, the… And it’s… Are the customers at all, like, part of that process [00:07:00] too? let’s say you guys get tightly rolled out and it sounds like you’re doing kind of regional releases too. do you guys have, like, customer loyalty programs or something like that where they can try this stuff out too early or something like that?

Riccardo: Not yet. I think it’s a good idea. Uh-huh. It’s a bit complicated to- I bet. … act because we’re in Copenhagen.

Luke: Right. Right.

Riccardo: And we would love to have people there-

Luke: Yeah, yeah …

Riccardo: from the markets.

Luke: Yeah, absolutely.

Riccardo: And the, the setup that we have is that we do everything globalized from the, Copenhagen office.

Oh,

Luke: okay.

Riccardo: Interesting. So it’s a bit difficult. We don’t have any developer, like, in the clusters. Yeah. So it’s a bit difficult to do local focus groups-

Luke: Sure. Sure, sure, sure …

Riccardo: right, from Copenhagen,

Luke: so. Right. Right. Right. Absolutely. And I know Pandora sells emotional kind of, like, high consideration products with- Yeah

jewelries, which can be tied to memories and meaning and have all that sentimental value attached to it. How does that context shape how you design an agent’s tone and, decision-making versus a more, like, I don’t know, Amazon or utilitarian kind of, commodified e-commerce experience?

Riccardo: Yeah, I think that’s the biggest difference- Yeah

you will see, like, [00:08:00] among agents. Our agent is- different from the tone, from how it propose your products, and that was a straight requirement from business from day one. Okay.

Luke: Yeah.

Riccardo: We started, and it was mostly me, like, because I come from business. I was in marketing before. Mm-hmm. Always been a nerd, and I just moved-

with the nerds- Yeah. … which is a natural move. But yeah. But, um, yeah, so the, the requirement was that one. I tried to myself improve, like, the instruction and the tone and everything, but I saw I was not able to, honestly. Sure. So we hired a conversational, specialist.

Luke: Okay.

Riccardo: Interesting. She has, like, a PhD in literature but also a minor in AI.

Uh-huh. So she’s re- really into the semantic meaning of words and everything. So, she got the requirements from branding and business, and then she translated those into instructions, right? And we ended up having an agent that has a very defined persona-

Luke: Mm …

Riccardo: with a brand vocabulary, words that cannot use and should use.

Luke: Interesting.

Riccardo: So it will never say luxurious and great, but it will use, like, branded words.

Luke: Yeah, yeah.

Riccardo: [00:09:00] Yeah. And the interesting part is also that- Yeah … you can give, uh, personas referencing other, like, popular personas that exist in the human world.

Luke: Oh, okay, okay.

Riccardo: Often you hear the example of if you ask a LLM to quote Harry Potter, it will quote uh, word by word because in literature and in the LLM training, it’s so frequently there-

Luke: Right

Riccardo: that it can recall it from training.

Luke: Yeah. Interesting.

Riccardo: So the same happens for, like, characters of- Yeah … movies and stuff like that.

So we gave, like, a famous character of a movie that I cannot disclose. But, like, she, like, she’s gonna behave a bit like that character-

Luke: Interesting …

Riccardo: because it’s, like, naturally fitting with the Pandora brand voice.

Luke: Yeah, yeah, yeah. Yeah. And, and the audience is pretty, you know… y- you know what the segment is and the customers are, and I’m sure you can even adapt that, too, as you grow out into different- Yeah … things if you want. It’s super interesting. how do you think about the relationship between- classical personalizations and what agents can now do.

I mean, it’s somewhat related- Mm … to what we were just talking about, but-

Riccardo: Yeah. The funny part is that before the PM role, I was, [00:10:00] uh, leading the personalization team-

Luke: Oh, really? …

Riccardo: in, in marketing.

Luke: Interesting.

Riccardo: So for me, it was like a natural progression to- Right … something that it’s even more into personalization, right?

Yeah,

Luke: yeah.

Riccardo: Because what we do today is that we intercept, like, real-time needs of customers-

Luke: Mm-hmm …

Riccardo: and we try to meet them there. So basically, a customer ask, I don’t know, next week is my girlfriend’s, birthday. She love necklaces. and I can spend up to $300. I need something meaningful for her."

So that’s kind of like a typical query, so that’s the result that you get is, highly, emotional in terms of, like, reply from the agent. It’s gonna tell you, like, “Oh, that’s such a thoughtful, uh, uh, thinking. this necklace represents, like, uh, her style,” da, da, da. So it makes it, like, very personalized for you in the reply, but also the products that it picks.

Luke: That’s awesome.

Riccardo: So

Luke: it’s- It’s really cool. Yeah. It helps out the, uh, boyfriends in the world. Yeah,

Riccardo: yeah.

Luke: Who might not be the best at selecting these things. Uh, uh, maybe I might be speaking from experience there. Maybe not. it seems like there’s real, like, [00:11:00] kind of tension in agentic AI between giving agents enough autonomy to be useful and maintaining enough control to protect the brand and the customer.

How do you navigate that balance? I know you guys said that you have, like, strong founder or business side input, but, like, wh- where’s the tension? Like, how do you guys handle that tension when you’re building- Yeah … something like this?

Riccardo: Yeah. So we try to give, uh, limited, uh, capabilities- Mm-hmm … but very, like, vertical and precise.

Luke: Okay.

Riccardo: So for example, we’re giving, the retriever, capabilities or product recommendation. We’re giving FAQ answers. We’re giving, some, like, off-topic that are branded, so replies that are… Yeah, I don’t know. If you ask, “Can you teach me how to code Python?” Gemma will reply always in the same way.

We’re deterministic through NLP, right? Oh, okay.

Luke: Yeah.

Riccardo: So it will say, “Hey, I would love to, but I’m here to help finding the perfect piece of jewelry.”

Luke: Interesting. Yeah,

Riccardo: yeah, yeah. Right? So it’s gonna hammer you with that answer-

Luke: Yeah …

Riccardo: and it’s not gonna, you know, so- go somewhere else, and it’s by design. That’s important.

It needs to be by design, otherwise if you give the LLM the, you [00:12:00] know, ability to just have more wiggle room- Mm-hmm … sometimes it will say something weird.

Luke: Interesting.how are you thinking about this where, in a world where a user might have other agents too? Yeah. is this even in the thought process yet or- Yeah

like of having your agent working with other agents or- Yeah … augmenting that?

Riccardo: So this is the topic, that we’re talking about with the big partners- Okay … technology partners- Yeah … today.

Luke: Yeah.

Riccardo: Because it’s coming.

Luke: Right, right. Right? Right. Yeah. It was, I was interviewing a guy the other day- Yeah … who said the, uh, agents, uh, buying and selling cars, like a dealership agent and a user’s agent, like, and it’s just like-

Riccardo: Yeah

Luke: wild, you know, that they’re in there.

Riccardo: So everybody talks about this protocol agent to agent.

Luke: Yeah.

Riccardo: So the idea is that we’re preparing Gemma, not now, but like the idea is to- Sure … prepare Gemma to interact with, like, the OpenAIs or whatever it’s gonna be, right?

Luke: Yeah.

Riccardo: Kind of like the top LLM providers that gives you, like, the personal assistant.

Luke: Mm-hmm.

Riccardo: Maybe we’re gonna have local, uh, deploy the personal assistants. Right, right. It doesn’t matter.

Luke: Right.

Riccardo: Right? And we need to be there with, like, the inventory that it’s someplace, uh, the product description that is [00:13:00] agentic.

Luke: Yeah.

Riccardo: Right? So- Right,

Luke: right, right …

Riccardo: the right, the right structure of data-

Luke: Uh-huh

Riccardo: for agents to be talking to each other and make its move, basically.

Luke: Super interesting. Yeah, yeah, yeah. Yeah. and how does agentic AI shaping brand control from, from where you’re at? Yeah. A- is that concern you at all?

Riccardo: Yeah. I think in marketing people, it’s like, “Damn,” like- … this is bad.

Luke: Yeah,

Riccardo: yeah.

Because they’re like… I think we’re in a place where Pandora is a bit lu- is luxury- Yeah … but it’s affordable luxury.

Luke: Okay.

Riccardo: So we need to be in both places, meaning we need to be, uh, mass reaching.

Luke: Yeah.

Riccardo: So we need to be there in all the LLMs and be e- easy to find. But you also don’t wanna lose completely the control of the experience.

So the scenario where all the shopping is gonna come from, like, external providers of agents-

Luke: Yeah …

Riccardo: it’s a bit- concerning, I think.

Luke: Yeah

Riccardo: Because we create the content on our website- Mm-hmm … on our social media. It’s very beautiful. It’s very [00:14:00] curated. It’s for our customers.

Luke: Yeah, yeah.

Riccardo: So how do we do that in a future where most traffic is coming from external agents?

Luke: Right.

Riccardo: That’s a bit of a concern that we have.

Luke: Yeah.

Riccardo: Some other brands that are, like, high-end could potentially decide, “Okay, we are niche. We are not gonna optimize for external agents. We want our customer to come to us in our shops, in our website. We don’t care about the other agents.” Right. “If they want us, we have a strong brand, we are desirable-

Luke: Yeah

Riccardo: you can come to us.”

Luke: Yeah, yeah.

Riccardo: I think we are in between.

Luke: Yeah.

Riccardo: We have that, but we also wanna be very, like, for everybody.

Luke: Yeah, yeah.

That’s what’s cool about what you guys are doing, where you actually have something in production, right?

So we have been live since basically October-

Oh, wow …

Riccardo: in Australia.

Luke: Wow, you guys are early.

Riccardo: Yeah, yeah. That’s why Agentforce put us everywhere.

Luke: Okay, got

Riccardo: it.

There are… There were, like, billboards with our, uh, CEO-

Luke: Oh, really? …

Riccardo: in like a tube in London.

Luke: Nice,

Riccardo: nice. Right? So that, that big. We took a small risk, but also the agent was, like, really… I mean, it [00:15:00] didn’t have so many tools. Yeah So, it was working fine. We also have service. We, we, we didn’t speak about service agent, but it’s live since, like, uh, one and a half years.

Luke: Yeah, let’s talk about that. Yeah. So, so how does that work with what you’re doing?

Riccardo: So, that’s easier because it has, like, where is my order? It does, like, an API call, and it gets the order. If you give order ID and email, it replies to FAQs. Mm-hmm. So, that’s kind of, like, industry is already there-

Luke: Yeah …

Riccardo: and it’s deflecting around, like, 60%-

Luke: Yeah

Riccardo: of the contacts. Deflecting meaning, like, people talks to the agent, and the agent somewhat replies okay-

Luke: Mm-hmm …

Riccardo: so that the customer doesn’t have to call the, the people.

Luke: Right, right, right. I

Riccardo: see. I see. So, that’s the metric, right?

Luke: Nice, nice.

Riccardo: With that, with the shopping agent, which is a bit more like my passion and it is- Yeah

a tricky one.

Luke: Yeah.

Riccardo: We’re live since October, but it was like a box at the bottom right. Like, the, the, you know, out of the box Salesforce provided.

Luke: Right.

Riccardo: But we felt like we needed a, uh, true agentic experience. Yeah, yeah. Because Pandora is all about branding, and meaning, and, like, it needs to look beautiful.

Luke: Yeah.

Riccardo: And our [00:16:00] stakeholders were like, “This is not that nice.” And I was like, “I agree.” How, how- Right. So, we hired a bunch of frontenders- Yeah … and we built, like, a full-on React native application for Gemma. Only for Gemma.

Luke: How has that been? Yeah. Like, I mean, October, that’s, that’s a while ago. Like, um, were there good lessons learned?

Were there things that- Yeah … were unexpected, , that- Yeah … you didn’t expect to see but had to kind of respond to?

Riccardo: Yeah so, I mean, the customers at the beginning were asking a lot, also, uh, service- questions in the e-commerce- Interesting … which we didn’t expect- Yeah, yeah,

Luke: yeah, yeah …

Riccardo: for example, because that’s what they’re used to.

Yeah. Like, they’re used to the bots on the customer support pages.

Luke: Right.

Riccardo: And they want to ask, “Okay, where is my order?”

Luke: Yeah.

Riccardo: So, there was a bit of education to do.

Luke: I could see that.

Riccardo: Yeah. But we are… So the cool part is that we are connecting through agent to agent, internal agents.

Luke: Yeah.

Riccardo: both the agents together.

Uh-huh,

Luke: uh-huh.

Riccardo: So, if you ask support questions, it’s gonna call the other agent.

Luke: Uh-huh.

Riccardo: You, you’re not gonna notice, but it’s gonna call the other agent.

Luke: Interesting. That’s so cool.

Riccardo: Yeah. So, we, we do that, and then we learned [00:17:00] also that, yeah, uh, sometimes you need determinism. Because we left everything, like route calling, math, everything to LLM- Yeah,

Luke: yeah, yeah

and

Riccardo: to the same model as well.

Luke: Yeah.

Riccardo: And it did a bunch of, like, weird things. Once it started, a customer was frustrated with the, uh, click and collect service.

Luke: Yeah.

Riccardo: And he, he said like, “Fuck click and collect-

Luke: Yeah, yeah,

Riccardo: yeah … first entry.”

Luke: Yeah.

Riccardo: And the agent said, “If you’re thinking about self-harm-

Luke: Oh my gosh

Riccardo: reach out to this Australian line number.”

Luke: Wow. Wow, it’s wild. Yeah.

Riccardo: And we were like, “Ooh, maybe we need like something a bit more, you know-

Luke: Catered …

Riccardo: deterministic.” Yeah. It’s interesting,

Luke: yeah.

Riccardo: Deal with this thing and frustration and everything.

Luke: Yeah.

Riccardo: So we needed to add that, and we are adding that. We’re keep going adding that.

Uh, yeah, so- I love that

Luke: though …

Riccardo: it’s

Luke: been a journey. I mean, that’s like the pioneering part of what you guys are doing, right? Like- Yeah … putting it out in the world and kinda seeing- Yeah … how… And if we don’t do this, you don’t see it getting better and, and these things. I mean, we’re joking about it now, but like, you know, it’s important that we learn [00:18:00] these things.

And when stuff can go wrong, it can go pretty wrong, you know? Like-

Riccardo: Yeah.

Luke: Yeah.

Riccardo: Yeah.

Luke: And

Riccardo: that’s why when we launched, we launched 10% of

Luke: traffic. I see, I see, I see.

Riccardo: Then we raised it to 50%.

Luke: Yeah.

Riccardo: Then it stayed 50% for like four months.

Luke: Okay,

Riccardo: cool. So it is A/B tested.

Luke: Yeah, that’s what I was gonna ask too. Yeah. Like, are there any, like, as somebody that’s rolled something out like this in production, is there any pointers that you have for folks that might be, you know, considering doing something similar?

Riccardo: Yeah. I think, be- Happy enough with the product, but don’t wait to have a perfect product. But

Luke: great advice.

Riccardo: Right? Just, just go- Yeah … at some point. Yeah,

Luke: yeah.

Riccardo: Like, you know, you need to have your main use cases covered, do a lot of testing, but then at some point just launch it. Maybe, you know, at the beginning we, we have put something like, “Hey, you’re among the first people talking with me- Yeah

Gemma.”

Luke: Yeah, yeah,

Riccardo: yeah. So, you know, I’m still improving. So people is also, like, okay with it,

Luke: right?

Riccardo: Yeah. Yeah, yeah. Like, they, they know that there’s new agents everywhere.

Luke: Yeah, yeah.

Riccardo: So ju- just go with it, but make sure that you first read the conversations. I remember was, like, [00:19:00] November, I went with my girlfriend to Seoul, and in the flight I spent, like, seven hours-

offline reading freaking conversations, right? But it tells you a lot.

Luke: Oh, yeah. Absolutely. It tells you a lot. Absolutely. Yeah. You got it. It’s like just getting good feedback, like, and, , I’m hearing you talking about how you’re getting service questions, and it just takes me back to, like, we’ll, we’ll go post something on Twitter and someone will report, like, some browser issue that they have with.

Yeah. And I’m like, we’re talking about a podcast episode we’re launching. You know, it’s, it’s just funny how, how people just naturally kind of, like, find the box and say something in it. yeah, I think, like, for folks that are interested in this topic, too, what resources were helpful for you, digging into- Yeah

this process? What ones do you recommend? And then where can people find out more about what you’re up to, where you’re putting things out?

Riccardo: Yeah. I try to be, like… I think the algorithm found out that, I’m in this bubble, so I got exposed to a bunch of, AI influencers and everything. So my recommendation is to look always, like, around.

Luke: Mm-hmm.

Riccardo: Also, I’m Italian, so I follow do- two Italian dudes.

Luke: No. Okay, cool.

Riccardo: not because they’re speaking Italian, but just because they are very famous. Right. One is, uh, [00:20:00] Antonio Gulli.

Luke: Okay.

Riccardo: And he wrote a agenting design patterns book.

Luke: Oh, cool, cool.

Riccardo: So I recommend to read that. Yeah. He, this guy works in, uh, Google, he’s, like, a senior director in CTO office-

Luke: Nice, nice

Riccardo: person, so he’s very, he’s very good. He posts on LinkedIn a lot.

Luke: Excellent.

Riccardo: So you can follow him. And then, uh, do you guys know Redis? I think all of us know it. Yeah. Yeah,

Luke: we all know it.

Riccardo: Yeah. The found- the founder of Redis- Yeah … is a Sicilian guy.

Luke: Yeah, yeah, yeah.

Riccardo: Salvatore Sanfilippo.

Luke: Yeah, yeah.

Riccardo: He has a YouTube channel.

Luke: Oh, excellent.

Riccardo: He, he has that in Italian, but of course he publishes a lot of stuff-

Luke: Mm-hmm …

Riccardo: on Hacker News and stuff like that.

Luke: Yeah.

Riccardo: He’s trying to be at the edge of, like, LLM. He does LLM training. Nice. He’s, like- Deploying models locally, it tries all the new models. Nice. Now, uh, Anthropic just dropped a new one and it’s, like, trying it.

And it tells you, like, immediately, “Okay, this model is good for this- Yeah … this model is good for that.” Yeah. So I think I like to hear from the people that is like top top, right?

Luke: Yeah, yeah. And where, where are you putting things out in the world if people wanna follow along with what you guys are doing?

Riccardo: Yeah, so in [00:21:00] Pandora we’re a bit shy with, like… We are not a tech company.

Luke: Yeah, sure.

Riccardo: So we’re- we don’t have, like, a tech, uh, newsletter or podcast or anything like that. Yeah. I think I will post on my profile, so if you wanna follow. Yeah, yeah.

Luke: Where, where can people find your profile, like, I guess?

Riccardo: Yeah, so on LinkedIn I think I will start posting a bit more.

Luke: Nice,

Riccardo: nice. So Riccardo Arnaldi on LinkedIn, you can find me there-

Luke: Awesome, man …

Riccardo: if you wanna follow, of course.

Luke: Great, man. Well, well, Riccardo, I really appreciate you making the time. Uh, this was a great conversation. Of course. I love it. I’ve learned a lot and I hope people go check it out too, especially our friends in Australia, right?

Yeah, I mean, at

Riccardo: the end of the month, uh, go to the Pandora UK-

Luke: Okay,

Riccardo: cool … and you will find the agent live- Excellent … so you can try it out.

Luke: Right on, man. Yeah- Yeah … I’d love to have you back, too, to check in on how things are going.

Riccardo: Thank you, mate.

Luke: All right, cool. Thanks

Thanks for listening to the Brave Technologist Podcast. To never miss an episode, make sure you hit follow in your podcast app. If you haven’t already made the switch to the Brave browser, you can download it for free today at brave.com and start using Brave Search, which enables you to search the web privately.

Brave also shields you from the ads, trackers, and other creepy stuff following you across the [00:22:00] web.