Transcript#

This transcript was generated automatically and may contain errors.

Hey there. Welcome to the Paws at Data Science Hangout. I'm Libby Herron, and this is a recording of our weekly community call that happens every Thursday at 12 p.m. U.S. Eastern Time. If you are not joining us live, you miss out on the amazing chat that's going on. So find the link in the description where you can add our call to your calendar and come hang out with the most supportive, friendly, and funny data community you'll ever experience.

Can't wait to see you there. We will go ahead and get started with introducing you to our featured leader today. And I will do a little bit more introduction than I usually do, because we're going to be talking with Claudio Rebelo. He is from Swiss Re, but the reason why we know who Claudio is, is because Claudio applied for an Impact Award. And I wanted to share a little bit about the winners. Let's see. Yes, Rachel's got a link in the chat for us for winners. Thank you so much, Rachel. So the sort of context of this is, I would love Claudio to introduce himself to us, but I'd also really like him to talk just a little bit about his application for his Impact Award and give us a little recap of that.

So Claudio, it is so wonderful to meet you. If you could tell us a little bit about what you do, what your role is, and then also a little bit about your application for your Impact Award.

Sure, will do. So first of all, thank you so much for having me. I'm really excited to be here. And also before anything else, just to say that any views and opinions I express here are my own and do not reflect one of my employer. And so my name is Claudio Ribello. I'm a model validation actuary at Swiss Re. So Swiss Re is one of the main providers of reinsurance and insurance around the world. And insurance is all about uncertainty and risks. So we have actuaries, they build actuarial models to quantify and manage these risks. And my job is to provide an independent assessment of these models. So in essence, I have to answer the question if these models are fit for purpose. So it's a second line of defense role.

So I've been at Swiss Re since 2012. And prior to this role, I was head of data analytics reserving of the commercial insurance unit of Swiss Re. And prior to that, I had actuarial roles in pricing and reserving as well. And my academic background is in economics and actuarial science. And outside of work, besides spending time with my family, I like to go to the library, play guitar and write music, or at least I try to. And then regarding the award, so I got the email for the impact award. So recently in the team, we've implemented a solution, which is how to make code easily accessible to our colleagues, including those who code and those who do not code. And the idea was to deploy your logic, your R functions into an API endpoint via R plumber, and then make those endpoints accessible also via Excel in kind of an architecture that makes it very, very easy to use your code and the user doesn't need to learn any pseudo syntax or handle or deal with messy user forms. So this allows code to be easily accessible to everybody.

into an API endpoint via R plumber, and then make those endpoints accessible also via Excel in kind of an architecture that makes it very, very easy to use your code and the user doesn't need to learn any pseudo syntax or handle or deal with messy user forms. So this allows code to be easily accessible to everybody.

Okay, and you won the impact award for the category of scale and reach. There were I think there were six different categories in the link that Rachel put in the chat. I highly encourage everybody to read through the stories of everybody that submitted and one and let yourself imagine what you could do at work with your suggestions and the things that you could sort of influence internally. Claudio, I would love to hear a little bit more about what drove you to do this. Like what were you seeing internally where you were like, okay, clearly we have smart people here who just are not coders the way that we are coders. What was going on?

The backstory: democratizing code access

So maybe like a big bit of a backstory and kind of how we ended up here. So around 2019, the Arturia function launched an initiative called Atelier with the goal to upscale the Arturial function. So basically in data science, programming, and communication. And this has had a huge, huge impact and has been a huge success and is sponsored by our group, Chief Actuary. And it's now not only just an Arturial function, but has an impact on in the whole organization. So that means that we have all these analytical capabilities, but the question is now how do we share these capabilities among our colleagues? So including those who code and those who do not code. And if somebody asks, well, if you code, then you should not have a problem. You can use the code easily, but you have a problem where let's say there's a really great R exclusive package, but you basically work with Python.

And there, this spec is no longer available to you. So you can then agree with your colleagues, so let's keep it simple. Let's all just code in R. We have everybody code in the same language, so we all can understand each other. So your colleague writes a script that can help you to do your work. And you open the script and you realize you can't make any sense of it. You don't know which functions to run first, what those arguments mean. You don't know if a variable is a data frame, a list, or a vector. So then either you spend a great deal trying to understand somebody else's code or you just give up and you rewrite your own script. So then you have a risk that you're not only duplicating efforts across teams, but sometimes you also end up duplicating efforts within the same team. So we end up basically splitting the expertise instead of leveraging it. So this is kind of the main problem that I've seen. And this solution is a way to kind of counter it. And that's why the way to go is to centralize your logic into an API. And that means you can have R and SAS, for example, calling an endpoint written in Python.

So basically, later you can build a shiny app that calls that exactly same endpoint without having to duplicate that effort. So the idea is to bring the logic of your code where the user is already comfortable. And the place where most users are comfortable, whether we like it or not, is Excel. So Excel can make HTTP requests. And therefore, you can basically, you're no longer limited to the native capabilities of Excel. And it's installed in every machine. Every person, whether they like it or not, is opening Excel at least once a day. And basically, you can access exactly the same capabilities that you could do with any other application that can make an HTTP request. So that's kind of basically the idea.

And then if it becomes really successful, then you can then build a shiny app on that, because now you have driven adoption, people are using it. You can then build a shiny app calling the exact same endpoint without having to duplicate the code. If you want to update, you can update the API and any client is updated automatically. So you maintain only one version control, one endpoint, easy to maintain and easier to share across teams.

Addressing concerns from coders

Okay. Sorry, I didn't mean to cut you off. We probably have a little bit of a delay. We are opposite sides of the world. I wanted to hop into Slido questions. But I also wanted to address the inevitable comments in the chat, which are, oh, no, low code or no code, like people who are code first, that kind of makes them nervous. Um, could you talk a little bit about the people who do code, how this works from their end, when they are interfacing with the results of like a no code environment?

Yeah, so I think there are two kind of important aspects. So one of this is that Excel does not replace your coding preference. So if something is in an endpoint, and you want to code in R, you can call that endpoint from R itself. So nothing stops you from continuing your preferred workflow. And maybe somebody developed something in Python, that now it has deployed logic to an endpoint, and you can just continue your preferred method of working. So it's not taking anything away. On the other hand, if you have your own code, and you know that most people are not going to use your scripts, you're also less willing to share. But if you realize that, well, if I share these scripts, many people are going to use it, then you're also more willing to invest in your script, make it clearer, and be willing to share your work with others. So I think like if you prefer to code, the preference is that both that more and the more people use your code, the bigger the impact you have within the company. So I don't see that there's a conflict in that people prefer to code and don't prefer to code.

more willing to invest in your script, make it clearer, and be willing to share your work with others. So I think like if you prefer to code, the preference is that both that more and the more people use your code, the bigger the impact you have within the company. So I don't see that there's a conflict in that people prefer to code and don't prefer to code.

Awesome. Thank you for the clarification. I wanted to make sure that we all knew we were talking about increasing accessibility in both directions, instead of making that smaller. Okay. Now let's hop over to some Slido questions. I saw a question from Natasha, which might help us build some context around what you do and what your team does and what you're facilitating, the type of work that you're facilitating. So Natasha has a little asterisk in her question, and I will ask it for her. It says, could you talk a little bit more about how you analyze insurance models to determine if they are good? I'm curious what that involves. And I think that that's a great question. How do you know if an insurance model is good?

Oh, sorry. I lost you. I lost you for a second. Uh-oh. Can you still hear me? Yes. I can see internet unstable, but. Oh, no. Okay. Well, I'm going to put the question in the chat as well so that everybody can see it. And hopefully if you can look at the chat, you can see it as well. Natasha asks, could you talk a little bit more about how you analyze insurance models to determine if they are good? Yeah. So it really depends on the model that you're validating. But first you go through several stages. So it really depends. Sometimes models are very big and you only assess a certain part, but in general, you sit down with the model owners or the people who own the model. You ask them to give you all the documentation. So you study the models, you ask a lot of questions. So the first stage is like a research stage.

Once you have your ideas, you can either ask them to stress test, to see all the validations that the first line of defense has done. And then because you have done many of these validations, you have your own list of tests like parameters. What happens if you change this parameter? What happens if you change your model assumptions? If you use a different model, where do you get the data? Is the data fit for purpose? Are the risks you're analyzing material or are there any blind spots? So a model validation contains many, many different aspects, but the main piece is, is the design of the model reasonable? And are the parameters well calibrated? And then you also try to assess what is the governance around it, around these models. So who owns the parameters? How many times are they updated? Where's the source of the data? Have you validate this data? Have you done the right adjustments to the data in order to model these particular risks? So there's many components on how to do this.

Modernizing analytics in insurance

Wonderful. Thank you. Let me see here. I have a question from Noor and I don't see an asterisk, so I will ask Noor to ask that question live. I'm going to highlight it in Slido and I will also put it in the chat. Uh-oh, Noor, we can't hear you. Oh, there you are. Okay, now I can hear you. Picked the right time because I was about to put food in my mouth.

So how advanced, you would say, like coding computational or tech-wise would you say insurance or the actuarial field compared to other fields? I say this only because I was a math major and my best friend who's also a math major became an actuary. And she's like, yeah, I do everything in Excel. And I'm like, we have other tools in Excel. Why would you use Excel? So what are some of the challenges with dragging, not dragging, encouraging a field to move in a more modernized infrastructure and direction? Yeah, so Swiss Re really has invested in analytical capabilities in several years. So this initiative of Atelier really has transformed the company, which is almost unrecognizable today. Yes, there's still a lot of things that are done in Excel, but in fact, we have a lot of analytical capabilities even before that. So a lot of the things we do were either done in Java or in MATLAB, but this initiative has basically tried to promote open source code and make it much more accessible to, let's say, the average user.

So before it was a clear separation between the developers, the technical actuaries, and the more business-oriented actuaries. And now you have quite across the whole organization capability. So we do use advanced analytics, machine learning, regression models, and of course, the traditional actuarial models across the company. So a lot of the things were done in Excel, but also we have our flagship tools and those are not done in Excel. Those are proper business-grade applications where you have a clear set of developers and the scientists and actuaries that maintain those main flagship tools. The benefit now is that there's a lot of things

you do upstream and downstream that now you can also do using a coding environment. And this is exactly what brought me to this, let's say, capabilities, because a lot of the things we do is to report and communicate our results. So in 2019, actually, this is how I learned R via this initiative. Before that, I was doing a lot of VBA, Power Query, and Docs, and Power BI. And then via this initiative, I got in contact with R and I got the book R for Data Science. I read it cover to cover and there was no going back after that. It's like you found the light. And I was really impressed that you do not need to be like a data software engineer in order to automate reports or to create shiny apps or such a web-based applications.

So it really felt like a superpower, but still requires a lot of investment in order to do and work with these things. So the ability for Excel to be able to communicate with R and Python basically says that Excel no longer needs to be, let's say, the weak link in this architecture, because now you can basically create a ggplot from Excel, acquire a document from Excel. You can do Monte Carlo simulation in Excel, copy lists within Excel, and you're just calling any code that is written in R or Python. So that's why I'm really happy with the initiative. At the same time, people don't need to install any libraries. They don't need to handle, they need to choose an IDE. And also, maybe the architecture, I can maybe elaborate a bit on the architecture, also makes it very easy for the user to use this code. So I think this is kind of the key aspect, because I have seen this. It was not fully original idea. I've seen an Excel tool that basically is calling a library in the backend written in Java, and the whole tool has a bunch of user forms, which really adds a lot of friction. So if you want to maximize adoption, you really have to minimize the friction. And I also didn't want the user to learn pseudocode. So the way that it's implemented makes it super, super easy, while still having access to most of all of the capabilities that you could do, let's say, in R.

Empathy for non-coders

Wonderful. Thank you. And I think it's important to also build a little bit of empathy around the fact that not everybody is going to want to learn how to code as like a whole new skill set. Imagine that you've been in your field for a long time, you're a very smart and capable person. You're maybe an actuary, or you're a subject matter expert in risk of a different, in a different part of the organization. And you have and you have kids and a life and bills to pay, and you have very little free time. And then someone comes at you and says, Oh, I'm going to need you to learn this whole new skill set. What a relief it would be to be able to contribute your ideas and contribute to progress inside of your team and your organization in the tool that you already know how to use. Because investing in learning how to code is a lot. It's a very big ask for people.

What a relief it would be to be able to contribute your ideas and contribute to progress inside of your team and your organization in the tool that you already know how to use. Because investing in learning how to code is a lot. It's a very big ask for people.

So keeping empathy in your heart for everybody is very, very important. And Claudio, I love this. I wanted to call out a couple of things that are happening in the chat, because there is a whole conversation around agent skills. Lots of questions around what skills are lots of resources that are shared around skills. For example, we had a whole data science lab episode, where we were joined by Eric Aiden Bowie, who described skills for us. Isabella has also given a talk at our pharma on skills as well. And at Posit, inside of the Posit repo, inside of one of the Posit repos, we have a whole skill repo on GitHub, full of skills from Posit developers. And those are, I think they're available automatically to Posit Assistant as well. So if you're using Posit Assistant, you have access to all of these or the model does.

Isabella, do you want to unmute for just a minute and talk about that? So we can all get on the same page. And then maybe I can extend another question to Claudio from Slido that's a little bit about AI. Sure. Yeah, I think you summarized it extremely well. Skills are instructions, an instruction or a set of instructions, usually a skill.md, which is a markdown file. And skills are very helpful if you are giving your AI agents some sort of repeated instruction. So the example that I like is, for example, there is a shiny BSLib skill. Say you make a lot of shiny apps, instead of every single time telling your agent, please create me a shiny app using BSLib. If you just write a BSLib skill, the AI will know, based on the description of your skill.md file, whether or not to invoke or load the skill. And then it will follow whatever instructions you have written out in that full set of instructions. And so that way, it's just another way of kind of like not repeating yourself with AI.

Yeah, it's the dry pence principle applied to utilizing an AI model. You don't want to repeat yourself every single time. And then Tamalu had said in the chat that it is a good example is maybe you prefer a particular style of UI. And every time you want your agent to do something for you, you want it to use that style. This is a way of avoiding having to detail that every single time. I want this particular style. I want things to look like this. It can just know that in the background for you with skills. So there are links to all those resources. And that lets me open up to ask some questions to Claudio about LLMs and AI and everything that are popping up in the Slido.

AI and LLMs in the actuarial field

So if you are up to that, Claudio, let's go over to a broad question that I see from Rizwan. If Rizwan is available to ask that live, if we could get them unmuted. Let's see. So maybe if I could, maybe before I forget, I have two colleagues who are also joining. Oh, yeah, yeah. So I have David Zadahi. He has been basically driving the internal promotion of the tool and is also helping me developing the R package and the source control. And he has had a lot of experience with providing an R solution in the Excel-driven team many years ago, which had clearly a lot of added value, but it was met with a lot of skepticism. So having a good tool is one part, but convincing people to adopt it is a whole other challenge. If any question for this topic, so David is kind enough to answer. And then I also have a colleague I used to work with in the Analytics team. His name is Sam Sekarsky. He was my hackathon partner, and he built an LLM workflow with the tool within Excel where the LLM, so he programmed with Elmer, trained the LLM to fill templates so you can give it a bunch of templates. It will learn based on the input and the output of your templates. And the training data then is saved in pins so you can then have the LLM fill the templates, the Excel templates for yourself inside Excel. So he is also on the call if any questions on this topic, he's also happy to discuss about it.

Hi, Claudio. So my question here is we have a mandate to integrate more and more LLM usage. We have an OpenAI and Cloud account. So there is a push from leadership to maximally utilize the token and try to convert all our data science process into an LLM driven process. Is there any push from similar experience you guys have? Yeah, but it's clearly, it's funny because when I was head of analytics, we started a discussion where other people should have co-pilot licenses, yes or no, that's really the beginning. And so this was a discussion, of course, I was advocating for, but this was kind of, oh, there's a cost and is it worth it? And then my main argument was, well, it's still very early, but we don't want to be, if it's something that's really going to take off, we don't want to lag behind. And funny enough, now, of course, the whole conversation has shifted around. So we live in a completely different world and now, yes, there's a lot of, let's say, strong encouragement to use AI and to implement LLM driven processes, right?

And to talk about the Impact Award and also Sam, when he kind of used the LLM in order to help your Excel, I actually working on a prototype where I have basically a process, an LLM that can help you to do, let's say, apply a bunch of reserving models within Excel. And then this LLM can, in the end, create a PowerPoint with the officer down package that creates a bunch of stochastic analysis on your triangles and then tries to also not help, not only help to interpret the diagnostics, but also help you to select the model. So yeah, there's a lot of push and a lot of exciting things happening with LLMs and AIs within the company.

Got it. Awesome. Thank you for your question, Rizwan. Okay. We had another one that was from Zach and has an asterisk. So I'm going to ask it for him so we can just stay on the topic of AI for just a second more. Zach asks, how much has AI influenced the field in general, the insurance or actuarial field in general? And are you reluctant at all to use it for sensitive data? Because insurance data is people data. Yeah. So first, insurance is a highly regulated industry. And for example, there are legal restrictions. So certain variables cannot be used, for example, for pricing or to discriminate in pricing. So everything is highly regulated and there's a lot of emphasis in being able to explain and justify your models and your results also to the regulators.

So there is one part of the AI stripping, let's say, the pricing and the reserving and the financial aspect of the company. There's another part that AI is used to help you summarize your reports, produce your exhibits, write a code for you that you're still responsible, of course, to make sure that it's correct. So it's a bit trick, but you get a lot of submissions, a lot of data, a lot of big contracts. And in the industry in general, there's a lot of new tools that can basically scan all these documents and look for keywords and exclusions. So it's slowly but surely starting and creeping in in different areas. But there's still a lot of, let's say, a lot of upstream processes. But once you get really to the final stretch, then it's still quite a sensitive topic. But yeah, it is really almost scary how fast things are progressing. And the conversations that we're having today were almost like science fiction just a couple of years ago. But a lot is happening. There's a lot of initiative, a lot of projects that are really driven by AI and implementation of AI. And the most important about what I enjoy, what I think is at least partially reassuring, is that the narrative is to augment the human capabilities, not to replace it. So this is the main angle that things are being done.

the narrative is to augment the human capabilities, not to replace it.

Awesome. Thank you, Claudio. I don't know if we want to have David or Sam chime in at all on this, but if either of you have extra thoughts to share around this topic about the integration of AI or push for AI or the good, the bad, any comments, feel free to hop in. No need to let us know in the chat. You could just unmute. Okay. I'll just give you a chance.

A day in the life of model validation

Let's hop up over to Slido. And I would love to, I know we're at the half hour mark here, but get a little bit more context around, Claudio, your role from Natasha. Natasha asks, what does a typical day or week look like for you? What areas do you split your time between? Are you mostly programming? Are you in more meetings? What does that look like for you?

So it really depends. My main role is model validation, but I'm very driven by data science. So I always find side projects to work on data science. But in terms of model validation, depends where I am in my validation. If it's in the beginning, I'm trying to set meetings, trying to find who's the owners of certain models. Then once that is set, trying to understand how the models work, read a lot of documentation. And then coding is quite important and helpful because you can, in certain cases, it's not, the point is that all validations are very different from each other. So it's not. communications are different from each other. It's not a recipe that you can apply. You can't apply the same recipe to all the models you validate but you can replicate, in certain instances, pieces of the model and that allows you to be able to code and to replicate and to test and to communicate, building visuals, ggplot, or plotly. This is really key. And you spend a lot of time also writing the report. Here, it's interesting because I write these reports before AI so you have to really write every word for itself and it's not something I do naturally but I realize it's how you really think and deepen your thoughts when

you're writing something and typing and try to explain because in the end you have to deliver this report to the main stakeholders and you need to be very careful and thoughtful how you present everything. It has to be clear but still detailed enough. Of course, AI is really, really helpful in writing and expressing your thoughts but I still try to word as much as possible myself because this is where a lot of new ideas or things that I didn't find while I was doing the review but when I'm writing and trying to explain it to others, you really get new insights. Writing the report is a part where I still invest a lot of time even though AI has made it much, much easier to do.

Well, this allows us to follow right into Amrit's question. Let's see if we can get Amrit to unmute and ask. I see him. He's in the camera on crew today. Hey, Amrit. I will put your question in the chat as well.

Thank you for the opportunity and welcome, Claudio. My question was I am from a fintech space and I don't know much about pricing or reserving projects. I was just wondering how do you get it implemented end-to-end? I think you answered a part of that question in our conversations now but yeah, and maybe an extension to that is how do you present it to senior executives? Maybe what gets them excited about the work you have done?

Explaining insurance modeling

I can't really go into too much detail when it comes to Swissy but generally speaking, insurance companies, they have what they call their flagship applications and there you still have clear separation between developers, business owners, so there's a whole kind of traditional way even before you add AI. The coding allows you then because these are kind of for the, let's say, the status quo but sometimes you have very peculiar types of policies or things that you need to review that can't be addressed by these, let's say, flagship tools. That's kind of where you can't have a one-size-fits-all.

In terms of how these models are created, it really depends on the line of business. For example, life and health is very different from property and casualty. Within property and casualty, things are also very different and it also depends how much information you have. If you have a lot of data, then you can apply much more sophisticated statistical methods. If you have less data, then you require a lot of expert judgment and monitoring and you really require a lot of underwriting expertise to price a lot of these deals.

You need this reserving function because once you sell the policy, you still know how much that has cost. First, you sell the product and only then it will cost you money in the way of claims. Therefore, you need to have the reserving function to ensure there are enough funds available in order for the company to be able to settle all these claims of the policy that has been written until now. It's very, very broad. It's very hard to synthesize in an answer, but it's a very fascinating field. Of course, statistics and math play a huge role. A lot of these are statistical-driven models. There you have GLM and then you have also live models. Of course, there's machine learning, but there's also a lot of underwriting expertise and judgment and monitoring and being very close to the risk. We also have engineering departments that can go for really the big policies. They can inspect the site and make also an appraisal of these risks. It's a whole world with many different pieces.

It is. I was just going to say, I come from insurance. I come from a totally different side than Claudio does because I came from P&C, which is property and casualty. I think that Claudio is in more commercial. Then we have Ryan here in the chat today, who is a CAT modeler and that is catastrophe modeling. How much might it cost us if a hurricane hits this spot? There are all kinds of different modeling. Claudio mentioned GLM. So many of the models that you see in insurance are GLM. That's a generalized linear model. If you are a person who wants to get into insurance, that is a great place to focus. It's like looking at the different types of GLM models, but it's really, really broad. There's so many different pockets and then there's reinsurance and there are, yes, it's very, very complex. Find somebody in insurance who wants to talk to you. Actually, I'm from P&C, right? It's so different that I would be completely useless in life and health. Yes, I would be completely useless in life and health as well, but also I came from auto insurance and I would be useless in homeowner's insurance, for example. So I know a lot about car insurance, auto insurance, and personal property insurance, which is you could insure your computer, your phone, you could insure jewelry that you're wearing, right? Like my wedding ring is insured, my computer is insured, all kinds of stuff. Find somebody who works in insurance and let them talk your ear off.

And one thing that Swissy does operates business to business, so we don't do personal lines. So we are an insurance company, so we provide insurance to other insurance companies and we also have a commercial insurance business unit, besides life and health. Hopefully that explains why I don't know much about what Claudio does and he might not know much about what I do. Okay, so let's hop back over to Slido. We'll get off of our very, very specific what is insurance tangent and let's go grab another question. Amrit, thank you so much for your question. I wanted to address an anonymous question because I think it is very funny. It says, hey, you pitched your API idea to Posit and you won an impact award for it. Have you considered pitching it to Microsoft because it uses Excel?

Pitching to Microsoft and side projects

Yeah, well, I thought about it and actually I have other things. I have a bunch of add-ins which I've developed throughout the years in old school VBA on how to manage colors, how to put the Brewer color package into my Excel, so I can apply the Brewer color package or the Veritas package to all my plots just in one go. So I have a bunch of different solutions and in fact, I will deploy with my API a bunch of, you know, own custom made, because the idea about this package, of this API, it works as an add-in. So this is really, really key. So I didn't want to have the API to be limited to a specific file, because then this file gets spread around and you lose track and then it's chaos again. So the implementation is really as an add-in, so it's part of your Excel application. So yeah, I could, because I think the application is, the implementation is really, really nice, but, and I thought about it already before, but I have no idea where to start. I don't know how I would even contact Microsoft and pitch this idea. I don't know either. Anybody out there in Hangoutland have connections at Microsoft or know how to do this? I used to have a friend who worked at Microsoft. They don't work there anymore, or a few of them. Okay, maybe somebody in the Hangout group can connect Claudio with somebody at Microsoft, because I think that that would be amazing.

And then you mentioned that you had some other things going on, and Noor asked a question, which was, what are some side projects, work or non-work related, that you are working on or excited about? I didn't know if you wanted to talk a little bit about any of your other little fun things you're doing. Yeah, well, this is, so one, just a completely kind of side project, which is this one to create an LLM driven reserving process and then, which is available in Excel, but it could be available in a Shiny app. So that's a nice thing as an API, where you can have all these kind of processes done with an LLM driven workflow. Another thing I would like to do, and this is kind of as a personal project, because of course, Posit Connect makes deployment very easy. All I need to do is to click one button and the API is deployed in Posit Connect, but I want to also deploy it privately in my own server. And I went to the workshop, so I went to the Posit Conf, I'm wearing the t-shirt, I don't know if you can see it, but I went to the tiny models workshop and I would like to translate all the tiny models into Excel, so people can do machine learning directly from Excel as well. So that's one of the side projects, which is not necessarily work related, but it's something I'm excited about.

Awesome. We love side projects. I would love anybody to share side projects in the chat, but also in the Discord server, we have a, like a show and share channel. And there's also a channel in Discord called only the silliest projects, where I highly encourage you to go put your wackadoodle projects, because they're so much fun. And then also go find, go search ham clock, literally anywhere in the Discord, define Hubert's ham clock, because that will, ham clock launcher, because that will make you laugh. All right. Thank you so much to Noor for asking our questions here.

Communicating with Excel-first stakeholders

I wanted to also ask, there was one anonymous question, which I felt fell sort of in line with Amrit's question of like, how are you communicating things to other people? Amrit was asking about executives. This anonymous question says, what are the most valuable packages that you use to share your work with those Excel first stakeholders? So this could be a broadened question of like, what is the best way to communicate with those Excel first stakeholders? And then also, are there any valuable R or Python packages that you have found in this sort of process of learning how to do that well?

So, there's one package that I really, really, really enjoy. And I think it's really powerful. And it's actually written by one of, developed by one of your colleagues at Posit. I forget his name, Richard Ione, or I- Rich Iannone. Iannone. So one of the kind of really, really, and maybe it's kind of generally overlooked, it's the point blank package. So I think that's a really powerful package for data validation. And when I was at Analytics, I developed a Shiny app where it's using point blank to validate data. And it's really, really powerful. And the syntax is so, so easy to learn. And so there's actually, you provided a webinar a couple of years ago, an R form, a webinar where point blank is presented. So I really use that a lot. I also use a lot what's called the officer package and office down in our markdown, because to automate reports, it's really good to have things to be knit in HTML. You have all the interactivity that's really great, but sometimes you cannot automate a hundred percent. So to knit things in PowerPoint in Word allows the final users to still edit or change the output. And then it becomes much easier for them to adjust anything that they need to adjust. So the officer package and office down package for our markdown is an extension of our markdown. It's really, really useful. And those are the packages that are probably overlooked.

I will hop in here and say all my votes for the office verse of packages as well. If you are used to interfacing with Microsoft products in a programmatic way, you know how hard it is. And Quarto is magical and does magical things. If you need to do some more complex custom things with styling for your word reports, Excel, whatever, the office versus packages by David Kohel are just fantastic. They are optimized for working with .RMD files. So with our markdown instead of .QMD files, please, please, please go check them out. If you have reached a limit with Quarto and Pandoc, definitely all my votes for that. Claudio, thank you for mentioning it. Also, Rich Iannone is going to be joining us next week, I think, at the Data Science Lab. We have something that we're calling Testtober. Isabella and I host the Data Science Lab on Tuesdays. We did R testing, testing in R this week. Next week, Rich Iannone is joining us live on Tuesday. If you are interested in that, come join us. He's going to be talking about testing and validating in Python. And just getting started. You don't have to be an expert. This is like dipping toes in the water type of stuff.

Okay. We had a question that I think it was from Brent because I saw it in the chat pass by me and then someone put it in anonymously in Slido. So I would love to ask it because it also is about communicating. So it says, how do you communicate the output details with no code individuals and what are some friction points in keeping source data or source models correct? Because I am just learning about the no code phrase. What a great question. Has a couple of parts to it. I'm going to put it in the chat here.

And hopefully we can kind of dissect it a little bit. So one point is how do you communicate the output details with a no code individual? And then what are some friction points? So how do you communicate the result of my work to no code individuals? Yes. And then the friction points was in keeping the source data or the source models correct.

And Brent, if you are up for elaborating or adding context to that.

Not really sure about the end part. But the beginning. So again, it's hard to go really too much detail. But when I write a report, it's always a little bit, how much detail do you want to go? Because a lot of this goes to senior management. And so the style that I use, this is just the style that I'm happy with, is that these reports sometimes are quite big. So I always make sure that the executive summary and the first part of the report is much more high level. A CRO or a chief group actuary reading the report, and they don't have time for all the nitty gritty details. But I also want to make my output and communication useful for the modelers. So I think this is how you also add value. It's not only about validating and pointing what is wrong, but it's also you want to highlight what is good. And also because you see so many different things around the company, it's almost like a privilege because you go from team to team and basically picking everybody's brain and picking up all the amazing work that people are doing. And then you can then also share that knowledge in other validations. So the beginning is kind of more high level, simplified. And then I'll say, if you want more details, please read these sections. So it's clear enough for senior management, but also later in the report, more detail for the model owners. And then really the super, super highly detailed, I put everything in the appendix. So my appendix tends to be quite large, but then I don't want them to feel to have to go midway through the report and just be completely exhausted to look at all the details.

I'm going to interrupt you really quickly because I saw Brent unmute, and I wonder if Brent might have a little bit more clarification on his question, because I'm also wondering if it is about collaborating with no-code teammates. Like if you have a subject matter expert who's like, knows a whole bunch about something, but isn't a coder themselves.

Yeah, exactly. Thank you for presenting and congratulations on the impact award. But yeah, I was more about how you communicate with the no-code individuals to then get that message across, to get that output to the stakeholders. Because for me, I deal with that situation a lot with other medical professionals who don't know how to code.

So it's a bit tricky because we work in a collaborative environment. So all my colleagues, other people I work together with, they're quite kind of experts in their field. And most people are coders. There are exceptions within the team who aren't coders, but they're extremely knowledgeable, or they code, for example, or they delve a bit into, let's say, other programming languages like MATLAB. So for me, it's almost like we cannot understand almost each other. But I basically work within the team with other experts in their field. So even if I present something that I did with R, they easily understand because we are all experts also. So it's also easy for my line manager. So because we are a second line of defense. So we are basically validating other people's work, and then our work basically goes to the experts and to the key stakeholders.

So in this case, I see what you mean. You have to communicate with a doctor who has no idea about statistics, has no idea about models, or has no idea about data. They have some. It's not that they don't. It's just they're used to doing with Excel, or they're not really sure what to use in a statistical test or model. And then trying just to communicate those ideas of what the data actually means. Yeah. In my professional career, not in my current role, but in my professional career, I had to deal with underwriters. And sometimes this is kind of maybe more analogous to your situation, but you have to communicate, for example, the output of a pricing study, or you update a benchmark parameter, and you have to communicate this to underwriters. And yeah, this is always kind of a little bit challenging, but luckily, at least for me, people I deal with are quite knowledgeable in the business.

And then I think as long as trying to avoid kind of very, let's say, fancy words, but also be clear, and honest, and open, and patient. But I see the struggle sometimes to communicate technical results to a non-technical audience. In that case, I think people are very visual, so I always try to communicate things visually. So for me, I love visual. I love visualizations. So when I did PositConf, one of the days was about visuals. It was very, very inspiring. I came back with lots of ideas. But I think communicating visually and clearly really, really helps to pass the message across. Wonderful. Thank you. Thank you, Brent, for asking your question.

Career advice

All right. We have three minutes left here. This has been a wonderful day, full of so many questions, so many thoughts that were probably boring for me on insurance. I'm so sorry, everybody. Claudio, I would love to ask you, in the next couple of minutes, if you could give us a piece of career advice. Is there something that you find really valuable, or that you wish you knew? What did you wish that you knew 20 years ago? And I'll put you, I'm just putting you directly on the spot.

In terms of advice, that's good. Yeah. Well, we talk about challenging the status quo, right? But the advice I would give is, sometimes you got out of your way to build something that you think will have a really positive impact. And you work on it. And in the end, you show what you've done. And then the adoption is not really what you expected or were hoping for. And this can be discouraging, because you put so much effort doing something that nobody requested. You ran out of your way. And then it didn't have the impact that you were hoping for. And this can happen. And my advice is, do not stop trying. Listen to the feedback. Try to understand why it didn't land as you were hoping for. Take those lessons. And always try again, because even a small positive change can make a difference. So that would be one advice.

And my advice is, do not stop trying. Listen to the feedback. Try to understand why it didn't land as you were hoping for. Take those lessons. And always try again, because even a small positive change can make a difference.

And maybe another advice is, now we're talking about AI and LLMs. And sometimes you feel like, is it worth learning a new skill when LLMs is just going to do that better than us? And I think that, although I question this myself, so it's not that I don't. We're all questioning this, yeah. But I really think it's important to keep learning and to keep building your expert knowledge. Because in the end, LLMs will never be accountable, or at least not within the foreseeable future. But there's always a person who's going to be accountable for what the LLM is producing. So there's always going to be a need for people who can validate, because I think it's going to be just a proliferation of black boxes built by LLMs. And a human needs to be able to validate and to be accountable for that. And I think in the end of the day, we are curious creatures. And I think we're always happy to know something today that we didn't know yesterday.

And I think in the end of the day, we are curious creatures. And I think we're always happy to know something today that we didn't know yesterday.

So I love it. Keep growing. Keep being curious. Don't give up on learning skills just because LLMs exist. And also keep trying. I learned today that the woman who wrote Lessons in Chemistry, if anyone has read the book or seen the show, got rejected like 98 times in a row for her novel and got accepted on that like 99th or 100th time. And I bet she's glad that she kept going for 100 times, right? Michael says, he loves that book. I haven't read it yet. Somebody needs to tell me to read it. It needs to go into my 2D red pile.

Okay, everybody, this was so much fun. We would love to see you next week at the Hangout. But Isabella and I would also love to hang out with you on Tuesday with Rich Iannone, creator of the GT package, creator of the point blank package. He's talking about testing in Python. Come hang out with us. It's so much fun. We chat on Discord, pos.it slash dslab if you would like to register. If you're not already, it's the same time slot as the Data Science Hangout. It is just the lab on Tuesdays. Okay, everybody, we are going to save the chat. Isabella just saved it for you if you'd like to download it. And I think she will also put it in the Data Science Hangout channel on the Discord server. I see a thumbs up. All right. If you'd like to save the chat yourself, there are three dots in the top right hand corner of your chat screen. Unless Zoom has updated and has changed that, hopefully that's still the same. Claudio, thank you so much for joining us today. I hope you had a wonderful time as well. Thanks for having me. Bye, everyone.