Transcript#
This transcript was generated automatically and may contain errors.
Hello, everyone, and thank you for joining our webinar. Today we're going to talk about two most shiny things. The first is obviously AI, but the second one is very special and unique, diamonds.
You will hear about Brilliant Earth AI transformation story, how they went from siloed data sources to unified AI powered analytics built on Snowflake and enabled by Posit.
But before we dive in, let me mention a few housekeeping items. First of all, please feel free to ask your questions in the Q&A box, and we will answer those questions at the end of the webinar. And if you experience any technical difficulties, please just refresh your browser and it should be fixed.
Okay, and one more thing, legal fine print. Don't rely on forward looking statements for your purchasing and investment decisions.
Speaker introductions
Let's start with speaker introductions. So my name is Sam Sergunyan. I lead industry marketing for retail and consumer goods here at Snowflake and thousands of retail and consumer goods companies are using Snowflake for their data and AI workflows, starting from collecting user profiles, predictive modeling, supply chain optimization, and many other use cases.
Joining me today is Matthias Mueller, senior director of analytics and data science at Brilliant Earth. Matthias, do you want to introduce yourself? Hi, thanks for having me. Yeah, like you said, my name is Matthias Mueller, I run analytics and some data science at Brilliant Earth. We're a fine jewelry and engagement ring company. We certainly deal with a lot of diamonds, like you said, so excited to be here and excited to kind of talk about our use cases and how we use Posit and Snowflake.
Thank you. Great to have you. And also with us today is Natalia Drobau, senior solutions advisor at Posit. Natalia, please introduce yourself. Hello, everyone. I'm Natalia Drobau. I'm the senior solutions advisor on the Posit team. I spent the past 10 years working with different analytical solutions, working with customers as Brilliant Earth to try to help them to find a way from their legacy tools into moving forward. Thank you. It's great to have all of you here.
The retail challenge landscape
Okay, let me set the stage why this conversation even matters. Well, companies face genuinely challenging environment right now. And those challenges could be grouped into three categories. The first one is around conscious consumers. According to Deloitte's research, 47% of modern consumers consider themselves as value seekers. And that means that attracting those consumers require deeper understanding of their needs, their specific behavior, predicting their churn, and then addressing it with more targeted, more personalized campaigns.
Luckily, we currently have access to abandoned consumer data and AI technology that allows us transition from the traditional retail marketing into more personalized, more value-based agentic marketing. The second large category is operations. Obviously, in the current geopolitical environment and tariff wars, there are lots of existing supply chains got disrupted. But again, with AI and data technology, we're getting access to the intelligence that allows us to build unified planning systems and monitor supply chains end-to-end, and also enable almost real-time optimization of the store operations.
And the last big bucket is agentic commerce. Agentic commerce promises the similar scale, the magnitude of change to what we saw back in 1990s with the emergence of internet and online shopping. And now lots of retailers, they're trying to figure out how to optimize their product catalogs and build more personalized offerings to win attention of those shopping agents.
And you can see that the challenges are very different, but the way we address them is very similar. We're using data and we're using AI technology to enable what we call agentic enterprise in retail, an organization where agents work alongside with humans to scale operations and to multiply productivity.
The danger of agents on messy data
On the other side, what we see a lot, organizations getting super excited about agents and they try to deploy them at scale. And it's one thing when you're using enterprise license for charge GPT and enable all your employees with productivity tools, like so they can create their slides or draft their emails or even put Python codes. That's a completely different thing when you are trying to embed those agents to autonomously drive your enterprise workflows.
Many leaders still believe that agents are smart enough, so they will figure out your process, they will find the right data and they will basically do the right thing. And this is very expensive way to fail. Agents they don't figure out the messy data architecture. They cannot understand, they will try hard, they will burn a lot of tokens, but they can still make mistakes.
Here is an example of the test I did myself. I asked an AI agent to assemble customer 360 profile using a raw bronze layer data and then compare it to already existing profile in the gold layer, clean data. So obviously it took about 10x more time for agent to rebuild, reconstruct this profile and it burned around 9x more tokens. But this is something that I expected.
Something that I didn't expect was that the result was actually different. Because I didn't define a specific parameter lifetime value for the customer, so agent when it calculated it, it subtracted all the return transactions from this lifetime value. And this is not what we've done for the gold layer. So I got a different result. And the lesson from this exercise is that if you run your agents on a raw unprepared data without the proper semantics and the business context, you will pretty much spend a lot of effort and money and scale your mess.
And the lesson from this exercise is that if you run your agents on a raw unprepared data without the proper semantics and the business context, you will pretty much spend a lot of effort and money and scale your mess.
Snowflake's recommended architecture
This is why at Snowflake we recommend the following architecture to build your agentic enterprise in retail. Before you get to this top action layer where your agents operate, you need to enable three layers underneath, starting with unified data layer that is open, interoperable, supports both structured and unstructured data. And then you enable governance layer, which doesn't only control access to data for humans and AI agents, but also helps control agents' costs. And then the context layer, the one that defines business rules across different use cases, across different functions, customer 360 and personalization for marketing, product catalog, JT Commerce for merchandising, supply chain, store operations, financial planning for different business functions.
Snowflake platform provides building blocks for all those layers and also powers our partners like Posit to build applications natively on Snowflake.
Brilliant Earth's story begins
Okay, now let's talk about Brilliant Earth's story. Natalia, Matias, please take it away. Thank you so much, Sam. So we've talked about the platform, what runs inside of Snowflake, how Posit team fits there with a native app. That covers what's possible. What we want to do now is to show you what it looks like when a real team uses it. We're going to follow Brilliant Earth, many of you know them as the fine jewelry company with a strong e-commerce business and a growing network of showrooms.
So Matias, their data science and insight team lead has been on this journey with us. And the best place to start it is not the architecture diagram necessary. It's where they began. So because of, I suspect a lot of you will find yourself in the same spot. So we wanted to follow the journey from where the issues arose and how the solution became available with Snowflake and Posit together.
The bottlenecks before transformation
So before any transformation, Brilliant Earth's bottleneck fell into two groups. One on the left side were the tools. One on the right side were the people and processes. The second group mattered as much because better tools as Posit and Snowflake would not have fixed the issues by itself. So let's start with the tools on the left side.
First of all, the issues were the siloed legacy tools. The data science happened in silos. Different teams used different numbers, different copy of the numbers, and there was not a single side of truth. So this is where the Snowflake came in. The reports were disagreed and someone had to spend time working out which one was right. A lot of the work was manual, right? So oftentimes, the work depended on one person's steps on one machine. This doesn't scale. This doesn't reproduce.
Also, a lot of the teams, you know, we all have our preferred tooling, especially when it comes to data scientists. So individual teams trying to fix those issues brought individual tools to solve those individual problems. Each made sense on its own, but when it brings together, they cannot align into one platform that the whole company can use. And then with AI, the issue arose where can we future-proof those solutions, right? It works today. Somehow we patch it together, but will it work in a year or in two?
So then those were the issues with the tools. But when it came to the people as well, there was a lot of the hand-built reporting. So the reports, weekly showroom and e-commerce reports were built by hand every week. The time passes. The report was ready by the end of the week. By the next week, the numbers were already not the ones that the leadership wanted to see or the ones that you can action upon. The deliverables were static. We're going to talk about those use cases later on.
But again, this didn't help to have the static HTML for the team that could just read it, but not ask any additional questions. But when the questions were asked, they were asked from the one central team serving every department. So however good that team was, and going back to Matthias, however good that team was, it's a ceiling. Every request became a prioritization debate, and someone always loses in this sense.
So putting these two columns together, you can see the real problem. The issue was not a shortage of data or talent. It's just putting all these things together and provide them to work in a reproducible way. With that, Matthias, do you have anything to add to the issues that I stated?
Oh, goodness. No, I think that summarizes it really well. I think we've had, and we still have, incredible talent and brilliant earth. Some of the best people I've ever worked with work here. And it was really an exercise in trying to bring people into the same ecosystem and the same sort of platform to really harness all of that talent. And I think we've been able to build something pretty incredible over the last few months, I would say. And all of those issues that you mentioned have definitely been heavily reduced, if not
Like you said, we're going to talk about a few of those. But yeah, holistically, I think this explains it decently well, yeah.
Thank you. I know sometimes it helps to just take a step back and for someone outsider to look at it and bring them all together. So from that standpoint, was there a specific project or a moment in time when you decided all of these little issues combined into the bigger problem that you cannot fix incrementally?
That's a great question. I would say, so I'm not sure if there was like a specific project that spurred all of this. I will say that I think with the advent of AI that we're all obviously in at this point, I think it became increasingly clear that the current path that we were on wasn't the one that I think would put us into the AI future that we wanted to be a part of.
We are very fast paced here. We do act on trends really rapidly. And we want the same to be true for our data science and our analytics. So it was very important to us to kind of have a platform that supported that. And I think we just weren't entirely confident or we became increasingly more inconfident, I should say, that the sort of AI solution that we had in place may not be that one that we feel like is going to take us there. So that's why we took off on this path.
Solutions: modernizing the analytics stack
So from the issues, let's move into solutions. This is, I think, the most exciting part.
So here's what Brilliant Earth set out to change. Again, they didn't bring a separate tool for each team. Showrooms, merchandising, digital growth, and finance now serve from one deployment running inside of Snowflake. And let me walk you through the three examples that Mathias shared with us.
So the first one, this is for the growth team. So remember, I've mentioned before those static multi-page marketing mix reports. Now it's a Bayesian model written in Python and executed in Snowflake with an interactive Streamlit app on Posit Connect in front of it. Now the VP of growth doesn't have to just read the results anymore, right? It's not static. They can plan the media mix themselves, combining this domain knowledge with the correct tools. And again, they can ask questions without waiting for the data science team to answer them for them.
So this one, this was the first use case that was shared. The second one was the showroom operations, right? So we learned that more than a hundred users moved into governed reporting now, right? Remember before they all shared this, they all used those different tools themselves. The number didn't match, but now was, again, in a controlled environment with the permissions inherited from Snowflake. Nobody has to rebuild the access rules and the second tool. It all can happen inside of Snowflake. They can utilize the same tooling, the same data, all within the control environment. And most of all, they can share it as well.
The third one, and this one answers the hand-built weekly reports as well. The last part of it. So the metrics are computed in code against Snowflake. So the numbers are reproducible, but then it goes into the code API that writes a narrative on top of them. And then the report is published in Posit Connect. Again, in all three use cases, and this is what amazed me the most about the whole story with Brilliant Earth, is that different departments, very similar problems, actually, very similar issues, they all found a solution in one way or another utilizing the same tools.
So I've mentioned three use cases. Mathias, do you want to add anything else to that? Are there any other areas where the Snowflake is used, Posit is used?
There are many. You know, I would say, holistically, it's just a very versatile system. And I think the reason why those are the sort of three examples that I shared with you when we first talked about making this webinar was because it was showcasing how we're using this sort of ecosystem of products in very different ways.
And you're absolutely right. So the marketing mix model, for example, that model already existed as a Bayesian model before, but it was sort of a more siloed process where a data scientist was in charge of running it and sort of pushing it out to the team, whereas now we've gotten like a fully interactive data product that the team can actually interact with and run scenarios with and all of those different things.
And on the other use cases, it's sort of the same thing. We're really trying to sort of figure out how we can take the sort of good work that we can do on the data science end and put that into the hands of the experts that we have all across the business. I mentioned this earlier, but it is very true that the talent here at this company is just incredible. And really kind of being able to enable those teams with great analytics products is just very invigorating. And it sort of has become a bit of a sort of self-starter to roll different data products out into different functions.
We're currently actually migrating the merchandising team as well as the finance team into some of their own use cases as well. So it's been very, very successful.
Business impact
From your perspective, what was the biggest business impact from many of the use cases that you've mentioned for the Brilliant Earth? Yeah, I think if you asked each individual team, they would probably point to their individual solution. So there is sort of a very nice feedback loop between the teams that we build products for and sort of us of how to optimize and go from an MVP into a more sort of prog-ready type product, and then kind of iterate on that.
I will say as a company as a whole, sort of the third use case out of the three here is probably the one where we've gotten quite a lot of great feedback. The idea of essentially having government reports that we can still trust, where we as data scientists, we take all of the math out of the report and we rely on writing code to actually do those computations for us. But then putting sort of a narrative spin on it by putting it into like an LLM API. That's been extremely helpful because then we can get weekly reports that are sort of easy to digest and sort of follow the voice that a specific person or team is using and essentially puts the report directly into that stakeholder's hands.
That's been extremely helpful because then we can get weekly reports that are sort of easy to digest and sort of follow the voice that a specific person or team is using and essentially puts the report directly into that stakeholder's hands.
I would say holistically, if I had to pick one, that would probably be the one that most, at least is the widest adopted across the company.
Right, that's perfect. And I love the story where each and every team has the deer, all the deer. If I asked the growth team, they would probably say the marketing mix one has been fundamentally life-changing, but it's nice because it really puts the sort of data science and analytics group into the idea of how can I build products? How can I actually put out a data product that is helpful to the use case of my end user?
And I think in traditional BI, that's a little bit different, right? Where you're sort of like constrained maybe to like the sort of existing formats. And now that we have sort of AI-assisted coding and we have this platform with Snowflake and Posit, the sort of opportunities of things that we can build had always been there in the past, but there are just so much more accessible to us to kind of like rapidly prototype.
We've gotten into the habit of doing quite a lot of disposable analytics because it's just so quick. It used to be that you had to put everything together and prop everything. But the idea of, say, having a shiny dashboard and putting that together, hosting in a Posit and connect to the Snowflake database, that is now so quick for me to put together that I can put that into production as a sort of disposable artifact. And I don't feel bad because I don't have to spend hours and hours and hours building it. It's just so quick.
That is amazing. This is truly inspirational for any of who's going to listen to this webinar. I love that. In the position that they need to do the change. Listen to their story. This is what waits for you at the end of the tunnel. I love it.
The marketing mix model use case
So with that, not that I have a particularly favorite team at Brilliant Earth, but we're going to move into one of the use cases. We're going to zoom in into the marketing use case now.
So what you're looking at right now is just the illustrative version of the channel investment planner. The numbers are made up, but the workflow is real. The workflow is what we were talking with Mathias before, right? So how it works now, say the VP of growth has an extra 200K to spend in Q4. They set the budget and the time horizon. They add what they know about the business. So as Mathias shared, they have brilliant people in Brilliant Earth. And for example, a VP of growth can cap the paid social at, let's say, 60K. So because that's the channel saturation he or she knows about, which then he or she can model and it doesn't need to go to the data science team and ask to rebuild the whole report, right? They can just click the run scenario and get a recommended allocation across the channels with projected incremental revenue and blended returns on investments.
Three things to make this work. First of all, engine and governments. The model is version controlled. As you can see on your screen, it says model 3.2 tag. It runs inside of Snowflake. It isn't on someone's laptop anymore. It also comes with self-service execution. The VPs can run the scenarios on their own. Nobody files a ticket and waits for the next cycle. And it also has seamless deployment, right? So it's a streamlined app. It's deployed in the Posit team native app. So it stays inside of Snowflake boundary with the same governance and everything else.
So with that, anything else you want to add to this particular scenario or talk about here? Again, the visuals you see are just to illustrate the scenario. Those are not real numbers. But the effect that this made on Brilliant Earth, if you can share anything. Yeah. Obviously, it's an extremely simplified view of it. The real thing has anything from response curves to all of those different things. If you've ever been working with a marketing mix model, we do expose pretty much the whole thing inside of this streamlined application as a sort of front-end piece to interacting with the model.
So our VP of growth and the folks that live inside of her organization, they can access that and go and look into performance, channel assignments, run scenarios like in the case of Brilliant Earth. predictions, channel assignments, run scenarios like in the case here above. But the point is that this just wasn't necessarily something that we had here before. You know, again, we've always had a model. You know, we always kind of sort of use that to optimize our business. But the difference is that in the past, that model essentially was provided as a sort of output. So like a PDF of recommendation of sorts. Whereas now we have this sort of front end piece that can interact with what lives in Snowflake.
And to your point, it's a version control piece that really allows my data scientists, which are incredible at working sort of at that piece. You know, they're allowed to sort of stay, you know, in the ecosystem that they can do their work in, i.e., like Snowflake, under the hood, do the model optimization. But then, you know, a lot of the business knowledge or sort of the ideas of running certain scenarios that sits with the business, which then means that, you know, with this
piece that I have on the front end side, I can put that into the hands of the business and have that group interact with the data and the model without needing to be in code or, you know, sort of doing any kind of deep dive analytics. They see everything in there. It's all modeled out.
And, you know, it's just really sort of like I think of as an eye opener of, you know, being able to kind of put some of that into the hands of the people that will actually end up making the decisions and running these types of scenarios.
This is amazing. Thank you for sharing.
So I got so inspired by this story that I decided to create a short demo of it.
It's, yeah, it's a little of a little of a prototype, right? So I just for our audience, it's not going to be showing you real data. I created the fictional company trying to mimic the same workflow with the synthetic data. So with that, as they say, the picture worth a thousand words. So we're going to move into the demo.
Demo: marketing mix model in Posit and Snowflake
Most marketing mix models end up as a slide deck or static document that's out of the date by the next quarter. In the next five minutes, I'll show you how with Posit and Snowflake, you can create one that, for example, VP of Growth can ask questions of any day of the week and it never leaves Snowflake. You can run the whole Posit team suite inside your own Snowflake account on Snowpark container services.
So you start the application, launch the app, and you can see three products in here. First one is Posit Workbench. This is where data scientists work in different ID such as RStudio, Positron, Jupyter or VS Code. We also have Posit Connect. This is where their work gets published as application reports and APIs for the business. And we have Package Manager that supplies trusted Python and R packages to both.
Everything runs next to your data and everyone signs in with their Snowflake identity. So first, let's step into the shoes of a data scientist. We'll go into Posit Workbench.
In here, we see all the available projects. We can start a new session and select the ID that we want to work in. We'll go into directly into one of the sessions. This is Positron, Posit's data science ID. I'm reading directly from the Snowflake. So I can open any of the tables, see the information about it. I can also jump into the exploration of my data if it's new to me.
I can even ask Posit Assistant to tell me a little bit about my Snowflake data. And powered by the Snowflake Cortex, it will run some analysis on it and provide me with some high-level overview of my data.
As a next step, I can go into developing a model based on my data in Snowflake. This is Bayesian Marketing Mix Model in PyMC. This model will carry models carry over when an ad keeps working for weeks after it airs and diminishing returns for each extra dollar in a channel is worth less. Uncertainty is building, so every answer comes with a range.
After developing and testing my model, I can go into registering it into Snowflake Model Registry. Again, very simple, as I'm utilizing the same credentials that I logged in my Snowflake credentials. Pushing the model into Snowflake Registry, we can take a look at it and see that it's there after we run the code. It's a Snowflake object governed with grants and versions. You can call it from SQL, and rolling back to last month's model is a single statement. But the person who owns the budget doesn't write SQL, so let's give them an app.
Let's go back to Positron. And we can find out that our application that our data scientist created lives in the same project. We can even pre-run it in Positron and see it before publishing directly in the IDE. Then we can go ahead and deploy it to Posit Connect with just one click of the button.
As soon as our deployment is successful, we can now switch the gears and see what the user interaction will look like with Posit Connect and the same application. Now I'm the VP of growth. I don't need to know how the model works, but I can see exactly which approved version of the model I'm running in here. And when the team publishes the version 3.3, the app picks it up automatically. No need to redeploy anything. I can also see that I'm running my application against the Snowflake data as well.
But in the user interface, I don't have to do any coding. I can combine my domain knowledge, setting up additional constraints in the application, and run the scenario against the same model that my data scientist published in Snowflake. I can see the results right away.
I can check where the money goes today and also plan the full future mix as well, setting up my quarterly budget, optimizing or resetting to my current plan. I can insert my own means and max right here as well. If I run the optimized with my constraints, I can also save my scenario and present it confidently to the CFO or whoever is required.
Saved scenarios are written back to Snowflake, so Finance sees the same numbers. It's built in Positron, governed in Snowflake, used on Connect, and it's one model and one version from the data scientist's code to the VP's decision.
It's built in Positron, governed in Snowflake, used on Connect, and it's one model and one version from the data scientist's code to the VP's decision.
What a great demo. Thank you, Natalia. And thank you, Matthias, for telling the story. Actually, we've got quite a few questions from our listeners. So let's start with the audience Q&A.
Audience Q&A
The first question. How long did it take you to move from your legacy BI system to Snowflake and Posit? And what would you do differently now?
Yeah, I think I have a short and a long answer to that. I think the short answer is that the actual migration process at least on the front end wasn't problematic at all. The showroom users, for example, that we referenced in one of the use cases, it probably took like two months to rebuild some of the things that they were used to and kind of host that in a net new system.
I will say the longer answer to that is that we're actually trying to be very sort of intentional about how we build the sort of future out in terms of like reporting and analytics and data science. So, in many cases, we just don't want to take what we have in our legacy tool and then the one-to-one copy it and roll it out in Posit and Snowflake. So, it's taken a little bit longer, but sort of the reason for that is that we're just trying to build something that we want to use two years from now rather than just essentially copying what we have and moving it over. That's sort of the front end piece.
Obviously, there is a lot of work that has to get done under the hood to essentially move data from how it ever was ETLs into the old tool, into Snowflake. That's probably another team than mine, which is this great, but that work is like near completed as well.
The short answer is it was surprising to me how easy and quick it's possible, but we're also trying to make sure we build it responsibly. So, we're evaluating what we have and I don't want to rebuild 1400 dashboards in Posit. I just want to figure out what are the things that someone would like to have, work with that end user or stakeholder, and then build those things instead.
Because I do believe that there is an opportunity or any of those migrations always pose themselves as opportunities to figure out how we can up-level, how we can make it better than what we have at the moment.
Yeah, thank you. That's a great question. Well, I wonder if that was asked by an engineer that was slogging in the data warehouse somewhere.
Yeah, but basically, as you said, it's not just a straight migration. You actually use this as an opportunity to review your current workflows, introduce lots of improvements. Yeah, that's true. And it's a bit of an issue because our existing tool is or was sort of a culmination of data warehouse and BI solution. So, it's more than a simple migration from sort of one BI tool into another BI tool, if that makes sense.
I think we have another question, which is similar to, you covered it a little bit, but is there a particular reason why I decided to move from the Posit side? I'm curious as well. When you move to the native app versus just a separate box, separate server deployment or in cloud deployment, what was the reason behind it? And did you regret that solution ever?
Okay, that's a fantastic question. So, for those of the listeners that don't know, but Posit obviously comes in, let's call it different flavors. I guess, Natalia, you could probably talk much better about it than I could, but from my point of view, there are different deployment methods for something like Posit Connect, for example, right? You can have it sit on like an AWS instance somewhere, or I believe there's the cloud version of it. But then there's this idea of essentially having it sit right within Snowflake as a sort of native application.
I've used Posit now at multiple companies, and this is the first time that I've had the opportunity of using Posit in concert with Snowflake. And when I found out, and when I researched this idea of having Posit deployed as a native app inside of Snowflake, what drew me to it was really the idea of governance.
what drew me to it was really the idea of governance. So, the beautiful thing that you get with deploying Posit inside of the ecosystem of Snowflake is that Snowflake still governs all of the access, all of the user roles, all of the sort of tables that someone can access.
So, the beautiful thing that you get with deploying Posit inside of the ecosystem of Snowflake is that Snowflake still governs all of the access, all of the user roles, all of the sort of tables that someone can access. Any of those things are effectively built in, and Posit Connect can infer that, or Posit Workbench, or whatever flavor or product you end up using.
And we are on a SSO provider that provisions accounts for us. So, essentially, the whole ecosystem for me as a data scientist is just taken care of. So, I don't have to worry about messing with permissions. My IT team loves it because, frankly, I can't mess with permissions. So, it's just sort of like a system that works really, really well.
If you're familiar with Posit Connect, there are a few downsides here and there, or a few things that maybe you're used to. So, for example, on AWS, you can go straight into the config file and change some things around. Those are the things you lose with Posit Connect on Snowflake. However, holistically, I have not regretted deploying it in this style for a minute. I would do it totally again like this. It's been a really great system.
Thank you, Matias. And thank you for highlighting the governance. As we talked earlier, this is the second critical layer which we need to build to enable the entire AI transformation. Yeah, and it's out of the box, which is so nice about it.
Great. I have one more question. How did you get business users to adopt new dashboards instead of asking for their old reports?
Well, I think ultimately it's about building something that someone perceives as an improvement over what they have. Yes, we're all creatures of habit,
and that is potentially sometimes a blocker to adopting something new.
But in the case of, for example, this cohort of 100 showroom users that we migrated from the BI tool that we have to essentially a more custom solution in Posit Connect, what they used to access is essentially static reports that update over time, tabular format, customer lists, performance, tabular data.
And that was working for them. And we essentially gave them a very similar thing on Posit Connect, except we wrapped it into a package called Query Chat, which essentially allows you to take this data, put it into a format of having the information just like they were used to, but also providing a chattable experience in addition to it that's governed enough that I feel comfortable rolling it out to people.
Where instead of having to click through 15 filters to get to the view they want, they can just say, hey, I'm looking for the last 14 days of bookings performance for my showroom X. And then it just filters the dashboards for them automatically. It has sort of, I would say, a limited use graphical ability, so it does a little bit of visualization. And that's sort of the simple fact of it.
More advanced, you can also use Cortex with it so it can access your semantic layers that you've built into Cortex as agents to kind of govern this sort of chattable experience as well. And those are great things.
But as an end user, the great feedback that we've had and the reason why I think the adoption actually worked in this case is because a showroom operations professional, they went from having a sort of, I almost want to say Excel-like BI experience to now having a sort of more advanced experience where they still get everything that they're used to. But in addition to that, they're also getting some of those more current features set out in the market that allow me to kind of improve the experience of interacting with data.
Trusting AI-generated insights
All right, and thank you for mentioning Query Chat. I think one of the questions that I have for you here is about trusting AI, right? So you told us about a setup, right? But the question is, how do you keep AI-generated dashboards honest?
Oh, goodness. I think my answer to that is that I work every day to try to make sure AI doesn't do any math by giving it as many guardrails as possible. Can you, in the end, control if everyone's following those procedures? No, which is an issue, but you can try to put guardrails around it.
I work every day to try to make sure AI doesn't do any math by giving it as many guardrails as possible.
What we've done here at Brilliant Earth or what sort of I've been advocating for is that if we think of Snowflake as a sort of like beating heart of our data structure, that there are different consumption pathways that a user can choose to essentially access that data. You know, there is a chattable experience, right? So you can go to Snowflake Intelligence. I think it's not called Snowflake Codework, I think. But you know, you can chat with data and Snowflake, and access the semantic layer and you're on your way.
We've wrapped that into like Cloud as well. So essentially, we have an ability to do from within Cloud, access the Snowflake semantic layer and chat with data. And we put governance around that. We write skills, we try to kind of fence it in. But in the end, it's still sort of, there is always a possibility, a very real possibility that what the end user is getting is not 100% the answer that you think they should be getting.
There's ways around it. You know, I know you guys just released a new package. They are the Commons package that kind of tries to work around that. But there's sort of, that's our consumption pathway. And if you think about it, My thought for this pathway as a whole is that like, that is sort of like an ad hoc pathway for quick and dirty analytics.
Whereas things that we put on Posit Connect, we generally try to like fortify it and make sure that LLM is not the only thing that we put on Posit Connect. We generally try to like fortify it and make sure that LLMs are not part of the pipeline of actually writing anything that sort of in the moment. Do we use AI to write SQL code? Do we use AI to write R or Python code? Absolutely. I think everyone does these days.
But, you know, what we put behind our work inside of things that we have on Posit, those are code-based. And I can verify, I can repeat, and I can essentially rebuild whatever is inside of there. And I can trust the code. And that's sort of like how I'm thinking about it at the moment.
And I can trust the code. And that's sort of like how I'm thinking about it at the moment.
Maybe next month is different. There's so much innovation and it's so quick that at times I think it's hard to keep up. So for the time being, that's sort of like where I am, but who knows what it's like in a month or two.
So what are you saying is don't be afraid to try, but proceed with caution, right? I think that's a good one. I mean, like if you think about the use case that we talked about with the AI weekly reports, right? Like we were trying to be intentional about like, yes, AI can summarize and, you know, it's really good at that, you know. But, you know, I'm still, you know, if anything, I want my functions to do the math. So can I maybe give AI a skill with a function? Yes, totally. Maybe that's a bit better than just flat out asking AI to do some computation for you without guardrails.
Okay, I see our time is up. So we have a few more questions, but we will answer them later over the email. Thank you very much for submitting your questions.
Thank you, Matthias, Natalia, for joining today's webinar. That was great. Thank you for all your insights. I had a lot of fun. Thank you so much for having me. I appreciate you guys. It's a pleasure talking to you. Thank you.
Thank you to our audience for joining our webinar today. If you want to learn more, please visit snowflake.com, posit.co, and feel free to reach out with any additional questions. And brilliantearth.com, if you're in the market for engagement rings or diamonds. Come visit us.
Absolutely. That's the most important and shiny things. Absolutely. That's the most important and shiny things. Absolutely. Thank you, Matthias. Thank you, Natalia. Have a good rest of your day. Bye.