Fiona (00:10)
This podcast episode is better viewed either on Spotify video or on YouTube and not Apple podcasts because we have some really interesting things to demo.
Sarah (00:23)
Welcome to unDUBBED the podcast that's unscripted, uncensored, and undeniably data. I'm Sarah.
Fiona (00:30)
And I'm Fi And today we are witnessing something extraordinary, a live demo of AI directly querying Tableau Cloud using nothing but plain English.
Analytics and Insight Specialist at PowerLink Australia, where he leads enterprise grade analytics for Queensland's high voltage energy infrastructure.
Sarah (00:55)
Darragh is a Tableau ambassador, a top 10 Iron Viz qualifier, and has over a decade of analytics experience across energy, commercial, property, international, education, and public service. He's worked all over the world, from New York to London, and even Ecuador. And today, he's joining us from Brisbane, Australia, where he's pioneering integration between Claude AI and Tableau through something called
Model Context Protocol
or MCP for short.
Fiona (01:29)
If you've ever dreamed of asking your data questions, like who are the top performers, and getting instant visualizations without learning SQL, without building dashboards, without waiting for analysts, you're about to see that future happen in real time.
Sarah (01:45)
So before we dive in, remember to like and subscribe to Undubbed. Most importantly, please share this episode with anyone who works in data because what we're about to show you could fundamentally change how we all interact with business intelligence tools. Welcome to the show, Darragh.
Darragh (02:06)
Hey there, Sarah and Fi. Thank you so much for the great introduction. I'm so happy to be here and talk about a topic that is lot of interest to myself and a bit cutting edge, I guess. So yeah, thank you.
Fiona (02:19)
Yeah, it's awesome to have you on here and I'm so grateful that you took the time out in your busy schedule this evening to help us out and to really show us the art of the possible with MCP.
Sarah (02:32)
Before we dive in to the demo, can you quickly explain what we're about to see? What is MCP and what should our listeners care about?
Darragh (02:41)
Yeah, thanks Sarah for that question. hopefully I don't stuff up this. I'm on that learning journey as well, but I have been doing quite a bit of reading and research around it and of course experimenting with it. So to give you a bit of background, Model Context Protocol or MCP, they call it the USB-C of AI. So it's essentially a universal protocol that
allows AI applications such as your chat GPTs or your Claude or any other AI applications built into other software like Visual Studio Code. It allows them to directly connect to basically any data source without doing a custom integration like coding up some kind of application to do the communication. It's just a standard protocol. I like to think of it as
a standardized gateway that lets models connect to and retrieve information from various external sources in pretty much real time as well. to give a bit more background about it, was invented by Anthropic who made Claude about six months ago now and it's quickly becoming the de facto standard for allowing AI systems to communicate with other types of data sources.
Fiona (03:55)
Yeah, I love it. But it almost sounds like we need to learn another language with all the acronyms in there of the USB-C of AI.
Darragh (03:59)
You
That's right.
I was trying to think of a metaphor today, how I could better explain, trying to explain it to myself, because you are right. It's kind of a lot of techno babble in some ways, you've got all this data out there, like a giant library of information or books sitting on a shelf, but imagine you had no catalog system where you could look up stuff to find it.
MCP is that catalog system. It's your way of being able to look up the catalog and then being able to find that information on the shelf.
Sarah (04:34)
And really interesting to hear it's only six months old as well. Yeah, amazing.
Darragh (04:35)
Yeah,
it's pretty new. I think there's, I saw a stat that there's something like 10,000 MCP servers at the moment.
Fiona (04:46)
I'm hoping now that you can share your screen and I'd love you to walk us through your setup what's running right now and what we should be looking for with how you're going to walk us through.
Darragh (04:58)
Yeah, so
you can see the screen, all good.
Fiona (05:00)
Yeah, and can
Darragh (05:00)
So this is Claude. Have you guys used Claude before? Yeah. So you're kind of familiar with how it all works. But I've set up a project, Dub Dub MCP just for this session. So I can give you a quick idea of how I can use it with using a Tableau Cloud example.
Did you need me to kind of go through and show you the specifics of the setup itself? Or are you happy just to see how it's integrated into Claude
Fiona (05:27)
So you don't have any project instructions. This is all just a standalone project. You haven't trained it on anything so far. Okay.
Darragh (05:35)
No, no, no. So
I've got the Tableau integration here. So what I essentially do is create like a specific JSON file and then integrate it into Claude, And you can see here, at the moment, Tableau have only opened up.
bunch of abilities. what happens is they've got an MCP server that allows you to connect to Tableau Cloud. And then you can run these queries on cloud from within Claude So I can do things like list data source, list fields, query data source, or read the metadata. That's all you can do at the moment. However, this still allows you to do some interesting things. So it's still very early days.
But I can show you how it works let's hope this works now. I did test it today. It worked fine, but Look up my tableau cloud and tell me how many data sources you can find
Now, it already knows because I've ticked that integration on, I've told Claude, yes, you can use Tableau Cloud. It essentially knows that it can query my Tableau Cloud instance. And when I set up my connection or integration, all I had to give it was the location of my Tableau Cloud instance.
I had to set up a personal access token on cloud and get the access key. And that was pretty much it. So it was literally four steps. you might have seen on the blog I wrote down the steps. All I had to do was put that in a JSON file and then integrate that with Claude. It was really, really easy. But you can see what's happened here. I've asked it to list the data sources. Now it's...
it's logged in to Tableau Cloud and go, I found 94 data sources on the site and started to give me a breakdown within Claude itself. And some of this is, here's some interesting data sources, multiple super store samples, some COVID related ones, some finance ones, some geographic and demographic data sources. So I'll keep doing a few queries.
I'll go, can you grab a superstore data source and tell me more about the fields in that data source.
What you'll notice is with Claude, usually tells you what it's actually doing. So it's using this listFields function. Remember at the start I talked about it, Tableau have enabled like five functions and that's one of them. So it's for some reason that it goes, can't find it using that one. So I'm going to read the metadata. And it just actually, without prompting, it just keeps trying to...
answer my question. And what you can see it goes, here's all the core fields. And I haven't even left Claude here. I haven't had to go off to Tableau Cloud and do this. It's just basically doing it for me. And since we're all Tableau people, we'll know that these are pretty much the common fields in Superstore. I think they're imprinted on the minds of anyone who's ever used Tableau before.
you can see it's gone in and started figuring out data types and stuff like that. And even looked at some of the advanced features. it's gone beyond what I've asked and go, here's like, here's the calculated fields. Here's some special bit and done this analysis of the data that's sitting on β a tablet cloud server.
you probably can imagine how useful this could be for a variety of reasons. you can do this analysis without having to manually go in. It's really great for building data dictionaries perhaps or something like that.
Sarah (09:17)
And reference checking, potentially different data sources and looking at how they're calculating.
Darragh (09:18)
Reference checking. Yeah,
so exactly. I mean, it's fairly limited at the moment in terms of using the Tableau one. I haven't actually experimented with other MCP servers.
I'm gonna query a data source that I did put up from my iron viz actually. Can you locate the hit machines data source?
So this is a data source that I used for when I did IronViz last year. (link is in the show notes) So I had to upload the data source to Tableau Cloud. You can see it has found it. I'm going to say, can you give me a list of top songs from 2022?
yeah, so this is a data source I do have access on. Now I know that those are correct because I remember doing this in the viz. so that's a very quick demonstration. I'm interested to see what you guys think on this initially.
Fiona (09:56)
Yay!
Sarah (10:12)
Well, I think it's very interesting, right? Like you think about the fact that you're just doing it in natural language and being able to query with what you've said in a setup. And we will share your setup blog and the show notes for those that want to set it up for themselves just through a basic API. And what did you say about four steps? if you're in the right environment, you've got access to your data and Claude, you could
Fiona (10:12)
it
Sarah (10:38)
really start doing a lot of analysis that would take a long time if you were using the native tools.
Darragh (10:46)
Exactly.
Fiona (10:48)
my mind is like going like the possibilities. Like I'm thinking, β do I now embed a Claude window that enables people to be in the dashboard at the same time?
Darragh (10:49)
The possibilities.
Fiona (11:04)
they can log in to Claude if the designer hasn't given them what they actually need to answer and just log in.
Darragh (11:11)
Mm. Mm.
Mm.
Fiona (11:14)
and query the data. Would that be possible, do you think, in an enterprise environment?
Darragh (11:20)
Interesting. Yeah, I guess so. I mean, I'm just trying to think of the security issues here. So, I mean, there's been a lot of talk around MCP and issues with security and authentication. I mean, I'm not over those debates, but I've just been noticing that people are going, how secure is this? So in terms of enterprise, I'm not sure if that's what you were asking, Fi, but...
the possibilities of, yeah, definitely if you have some AI application and many, enterprises do have that now, as long as you can connect it up via MCP to anything,
Fiona (11:51)
So I think that it's definitely possible, but there would need to be some big guard rails around it, like you say, from a security perspective. But the thing that really struck me was a long time ago, did a proof of concept on ThoughtSpot. And so we're talking six years ago.
Darragh (11:54)
Hmm.
Fiona (12:11)
So I'm sure the product has come a long way since then but at that stage they called it an Amazon like search and not a Google like search and what they meant by that is when people Google They would be more like help me blah blah blah blah. Whereas an Amazon you sort of say SHURE SM 7b microphone
Darragh (12:11)
Hmm.
Mm-hmm.
Hmm.
Fiona (12:34)
and it's much more tokenized in its own way and the way that it queries. And looking at the way that you interacted with Claude to get access into that data, it was seamless.
Darragh (12:37)
That's right. Yeah.
Hmm. Yeah. I mean, we're using Tableau Cloud as an example here, but you could have a Tableau Cloud going, you can also connect it up potentially to your data warehouse, or anything, like APIs as well, or just file systems and query them all at the same time. yeah.
Sarah (13:06)
Which would be amazing, right? You think of all
the organizations that have got so many siloed data sources out there and to be able to basically just throw access at everything and say, how many definitions do I have of customer churn? And get some answers back from all different departments would be fascinating.
Darragh (13:13)
Hmm.
Mmm. Yep.
Yeah, that's a really good point, Sarah. I mean, I was thinking along similar lines, for like a data quality or even governance point of views, like if you're looking at a dashboard or built in Tableau or whatever, and you could query the Tableau data source for this and tell me these values. And then also at the same time, query, you know, the tables in the database
that store the same information and is there a difference or are they the same? You could really delve deep into if you suspect there's issues in the data, it might have been transformed along the way You could use something like this to get some confidence. I mean, that's just one of many possible use cases. So yeah, it's an interesting space.
Fiona (14:09)
I'm looking at
these songs from 2022 and I've got Nirvana and I've got Fiona Apple, George Michael.
Darragh (14:12)
β yes. No,
Sarah (14:17)
George Michael?
Darragh (14:19)
I should explain some of the background this so actually, yeah, that's a really good point because these are clearly not from 2022. the data source I'm using it from is from my IronViz entry from last year, where I was comparing like the top 200 songs from the 90s.
They did two different rankings, the same magazine, 10 years apart, one in 2010 and one in 2022. And I was examining, how has taste changed? how do people think about music from that decade? How has that changed?
Sorry I didn't say it upfront. I should
Fiona (14:51)
that's
great because it's it's all about the context, right? And so the query that you've written, the first thing Sarah and I just did was, wow, and then we're looking at it going hang on a minute. That data doesn't quite feel right. Yeah. But having that context really helps. I just wanted to circle back to one of the other points that you were talking about, which was how
Darragh (15:00)
That doesn't look right. Yeah.
Fiona (15:17)
Essentially, you're taking the Tableau visualization out and replacing it with the query. In my mind, Tableau had already started to allow this or move towards it or not walk away from it. So for a very long time, we were unable to get into hyper files. We were unable to query the data in Tableau unless we were using Tableau. But
Darragh (15:21)
Mm-hmm.
Mm-hmm.
Mmm.
Fiona (15:44)
with Viz QL, they enabled a way for people to do that. So I think perhaps their roadmap and what they're seeing happen in the fabric of data, you know, the universe of data and where data sits, like you were saying, this could be
Darragh (15:49)
Mm-hmm.
Fiona (16:03)
the canvas that allows you to query all of those disparate data sources in one place.
Darragh (16:11)
Hmm.
Yeah, I would agree. It's interesting you talk about, VizQL there, you both probably have far more knowledge than me about how the Tableau backend works. But from what I understand when I'm querying Claude and we saw up here, it goes query data source.
that calls Tableau's MCP server, and then it runs VizQL after that.
I'm more than happy to do a few more queries if you guys want, unless you let's, β yes, go on. Yes.
Sarah (16:37)
Yeah, us see it.
Fiona (16:39)
While you're doing that,
just one thing. I feel like it's almost the DBT, the translation layer from a natural language query into whatever the source system is. I'm gonna make it up as I go along to try and narrow down on it.
Darragh (16:52)
Yeah, it's Yeah, I know look
We're still very early days in the MCP world. we could just like start, up our own language
Sarah (17:03)
up some acronyms ourselves.
Fiona (17:06)
I'll just get you to rerun that prompt for us and let's see what happens.
Darragh (17:10)
No worries. I was trying to do something a bit more interesting than just showing a list of songs. So this is a prompt I prepared earlier. I'm asking it, can you graph the top 20 songs in the 2022 list versus the top 20 in the 2010 list? we've got two different data sets here, one that ranks songs.
12 years apart essentially. So let's see what Claude does here. in this instance, we can't actually ask Tableau to do the charting within Tableau itself. So what it does is essentially goes to Tableau, grabs the data, like through the MCP server, and then brings it into Claude.
and then uses its own knowledge of how to chart in this instance, a combination of HTML and JavaScript to build graphics. Now, a bit of a spoiler alert, like it's working away there, writing that looks like CSS. The charting ability is probably not like enterprise grade, but it's not completely terrible.
it doesn't necessarily know what's the best chart to use, but it does provide some baseline functionality.
Fiona (18:22)
I saw just as the code was scrolling through padding in there, like it's going to have padding, which is one thing that so far we've been unable to fix automatically. And you would know from us working together for a
Darragh (18:25)
Mm-hmm.
Yeah.
Fiona (18:39)
at JLL, padding is one of the top things that I look for in great design. So it's great to see that
Darragh (18:45)
Yes.
Fiona (18:48)
it's including it in the code.
Darragh (18:51)
Yeah, so actually on the topic of padding, it's a really interesting one, Fi because there's one I've learned off, you being former colleagues and, the program that you ran in our old organization. But I'm convinced now that good padding is the key to good design. And this is why so many like, Power BI dashboards kind of look fairly bad sometimes, it's because you can't really do much padding in it. And that's where Tableau has a competitive advantage.
Sarah (19:15)
might say it's failed on this one.
Darragh (19:17)
It has failed in this instance. this is probably pretty meaningless. But the interesting thing about it is, see how it's got tool tips. It's just built them. it's gone on to the 2010 list of this pavement was number one with gold sounds, which is a very good song, classic song from the nineties. And then Mariah Carey is the top.
One in 2022, also a great song, fantasy, one of my favorites. But it's got tool tips in there. It's got like these, it's just built this legend, yes, not accessible possibly. And it's done probably what I wouldn't do, mapped position versus ranking. Thing about it down here is then just built like BANs out. I didn't ask her to do that. all I did was ask this one prompt.
Fiona (19:47)
Yeah.
Darragh (20:00)
and it's just queried the data source and then done all this stuff. And I built all these things, like these insights, mean, yes, the graph itself is not great, not great practice, but I'm like going, yeah, that's exactly right.
Sarah (20:11)
but it's only gonna get better, right? Like we've got to remember this is really
cutting edge, very, very new. And with hopefully with the right direction and instruction, it will get better.
Darragh (20:18)
Yeah.
we all acknowledge is not the greatest graph, but it's the potential here. I could go, well, Claude, that graph sucks. what would be a better way to graph this? just, you guys are viz Like, what are you thinking? Let's, I want to get your input.
Fiona (20:39)
when I was looking at this, was a bit confused by it because it seems like it's always the red dot on top and then the green dot below, can you hover over the first one for a minute? So you've got the Wu-Tang Clan there or pavement or whatever one, then go down to the one beneath it. they are two different.
Darragh (20:46)
Mm.
Yeah, it's
yeah, yeah, that one. Yeah, yeah, sorry. Wu Tang, there it is. And Missy
Fiona (21:00)
songs. I don't know how would I how would I chart it? How did you chart it for your IronViz
Darragh (21:06)
I
don't think actually I'd use some weird radial chart. I can bring it up on screen and show you, but maybe I'm asking the wrong kind of question. it's probably you can't compare it in a chart like this. Yeah. I mean, I can ask Claude, can you do a better one?
that? β
Fiona (21:23)
need to understand how the songs from 2010 were ranked in 2010 and how that moved into 2022 and likewise in the reverse.
Darragh (21:37)
Yeah.
Okay, so I need to understand how the songs from the 2020 list were ranked.
Fiona (21:46)
and how they moved.
Darragh (21:48)
they moved in the 2022 list, I'm assuming. Yeah.
Fiona (21:51)
ranking list,
and then vice versa.
Darragh (21:57)
Also, one, two, three, the opposite.
I'm gonna put vice versa in there as well. I might actually be a bit explicit here how the 2022 songs ranked in the 2010 list.
Okay, let's see what happens. Now, my prediction here is it'll completely stuff it up, but let's see what happens. Maybe Claude will impress me. I spent about four hours on Sunday morning just trying to recreate part of my iron vis. And inevitably kept stuffing it up, but
Fiona (22:16)
list. There we go.
Darragh (22:34)
it kept getting closer and closer the more like prescriptive my prompts were. So what you said there Fi the more detail you give it. And this goes for any, large language model or AI application, the more prescriptive your prompt engineering is, the better results you're going to get.
Sarah (22:48)
At any stage did you give it an image not necessarily your iron viz but any kind of image to help guide it?
Darragh (22:53)
Yes. β
Yeah, so I have and I can demo that too next if you wish. So I basically, did it a few ways and I've actually got these prepared if we want to explore it. I gave it the full image of the multi-part visualization built in Tableau public and go, can you just recreate, this from this data source? And then it did a okay job. It
recreated most of it, not using the charts I used, but it got the sections right. And then I went to looking at it one by one and trying to perfect the individual images. But then I ran out of Claude tokens. And once you hit that limit, you have to wait till the next day. So I didn't have a chance to keep going. So hopefully I don't run out of tokens here, but I should be right. I haven't been using it much today. But β it's done. OK, let's see what's going on here.
What do you want?
Fiona (23:46)
7 stayed in the top 20 when we were on from 2010 to 2022. That's really quick at that. It's got some nice BANs I'm not a fan necessarily of the colors that it chooses, you know, c'est la [vie] And then 13 have dropped out. Great. yeah, because there was 20. One moved down, six moved up. Interesting.
Darragh (23:50)
Yep. Mm-hmm.
And that's...
Yeah.
Fiona (24:12)
yeah, it's not done it in a chat, but done it in the table.
Darragh (24:16)
Yeah, that's interesting. β
Sarah (24:19)
which is probably
Fiona (24:19)
Interesting.
Sarah (24:20)
easier to read with such movement.
Fiona (24:22)
Yeah.
Darragh (24:23)
Yeah.
Fiona (24:25)
Yeah.
Darragh (24:26)
yeah, it's like what you would see on a song chart. everyone remembers looking at the top 20 ARIA songs It's like that. sure if you noticed, it allows me to, yeah, it is. It's just built interactivity into it. So it's done your second one. I just hit this chart. I mean, it's just built in JavaScript.
Sarah (24:30)
top 20s.
Yeah. And is it, yeah, is it interactive?
Fiona (24:36)
Yeah.
And then the slope chart,
slope chart, slope chart.
Sarah (24:44)
yeah, let's look at the slope chart. β interesting.
Darragh (24:44)
Yeah, and oh, let's have a look at that.
Fiona (24:48)
This
is really good.
Darragh (24:50)
This is pretty cool. And this is what I get excited by, β it's kind of exciting and also kind of scary in a way because it's my whole career
in the last 10 years has been about having that expertise to build these things and things like Tableau or Power BI, whatever. β perhaps you just skipped that bit now. You don't necessarily need to know how to build a radial chart or something like that because potentially the AI could do it or Claude or whatever.
Fiona (25:17)
Possibly, but there's a couple of things to unpack there. So first up, there's essentially a table version where you're reading things versus visualization where I can see both the movements, so both of the queries that I had. if I had to think off the top of my head what the best chart was, like you asked me.
Darragh (25:19)
Yeah.
Mm-hmm.
Hmm.
Fiona (25:42)
I struggled to conceptualize. I'm sure I would have come up with a slope chart eventually, but it's come up with us in a few minutes and visualized it. So that's my
Darragh (25:43)
Hmm.
eventually.
Hmm.
Fiona (25:55)
insight. My second insight was just around the tabular reading version, which is more cognitive load.
Darragh (25:56)
Hmm.
Mm-hmm.
Yeah.
Fiona (26:05)
So it's more cognitive load for me to be looking at this. There's a few things in terms of the alignment that isn't so great in here, but like you said, reminds us of something that we would think of with the chart. So visually you can read it, people who are less comfortable with data visualizations may even prefer it as well.
Darragh (26:32)
what do you think about that? β
Sarah (26:32)
I think it's really interesting.
I personally when I saw this viz my mind went straight to a combination of both. the the list of the 2010 on the left, and the slope in the middle and the list of the 2022 on the right, but reducing as much cognitive load as possible. So, I think the rank is really important.
Darragh (26:48)
Hmm.
Sarah (26:54)
If you're showing the slope, you probably don't need as many numbers as what the list is showing. And I think into that response around, our jobs and everything, and our jobs are transforming massively right now. But right now, I feel if anyone else looked at this, they'd probably get quite lost in it. And it still needs that refining and expertise.
Darragh (27:08)
Mm. Mm.
Yes, you do need that expertise. Yeah, you're 100 % right.
Yeah.
you made me think about it in a different way there, Sarah, when you were kind of talking about the combination.
Sarah (27:25)
Yeah.
Should we ask Claude and see what they come back? Like
Darragh (27:28)
Yeah,
so what kind of query do we want to ask?
Sarah (27:32)
Well, can we show, the tabula and the chart together?
Darragh (27:37)
and Tabula. I don't know if you guys use AI apps like me. I always compliment the AI, like when it does a good job. I don't know if other people do that. But I'm like, good work. Like it feels like this.
Sarah (27:45)
Mm.
Fiona (27:50)
Anthropic would hate you, they'd hate
you, because you're using up more juice just to get that token.
Darragh (27:55)
Yeah. β
Sarah (27:56)
I just
introduce it. I'm like, that's great. And then carry on.
Darragh (28:00)
That's right.
because I think of these these things as really hard-working interns like they can they can work really hard but they're kind of a bit dumb Over enthusiastic. That's a great particularly Claude Yeah, yeah, it says you're doing a good job and I go I don't think it's that good So can we show the slope chart and the tabular information on one
Sarah (28:04)
Yes.
Fiona (28:04)
Yes.
Sarah (28:07)
Over enthusiastic.
Fiona (28:12)
and over-complementary as well.
Darragh (28:22)
screen is that what you are asking Sarah? Yeah?
Sarah (28:24)
Yeah.
Yeah.
Fiona (28:27)
Well, this.
Darragh (28:27)
It's kind of sometimes a bit
of a roll of the dice. And one thing I haven't experimented is when I've got something I like, telling Claude that's what I want and then give it new data obviously you could go down this templating approach, like where you build up like a template using a data set, β much like, you know, many BI people do in their day to day use kind of like corporate.
templates to build up dashboards or visualizations and you go, here's new data, just do it like this. I haven't experimented it, but I'm sure it could do that. What I might do after it does this one is I can show you some of it using images like screenshots if you're interested.
Sarah (28:57)
Mm-mm.
Fiona (29:07)
for sure.
Darragh (29:09)
Anyway, hopefully you can see the screen here. β So if you remember, we asked it, can we show the slope chart and the tabular information on one screen? it went through and wrote a lot of JavaScript. One thing I'm not sure if you noticed when I was running this, it brings the data into the HTML JavaScript file. So it is static in that sense.
Fiona (29:12)
Yes.
Darragh (29:30)
It's not dynamic. And this is guess, big thing. I mean, I could just ask it to run more queries, but as you can imagine,
Sarah (29:33)
drawback.
Darragh (29:37)
It's going to take up a lot of bandwidth out of your AI application. so what we can see here, it's kind of done what you asked, Sarah. you guys can see it on the screen here. It's created some of these BANs it doesn't look great. I mean, it has put it on on the one screen. It's got the slope chart here, but then it's kind of just like compressed it over.
Sarah (29:59)
kind of smushed everything
Fiona (29:59)
Can
you, if you drag that left thing across, can you do that? Yeah. There we go.
Sarah (30:00)
together.
Yeah.
Darragh (30:04)
Yeah, let's see how that works. yeah, yeah.
This is okay. I guess on wider screens or wider, you know, displays. Is this acceptable? Yeah, I'm keen to understand what your gut feels on this. I basically did what you said, I would say.
Fiona (30:20)
I think when you have a full screen, it would be fine. What's that publish button?
Darragh (30:25)
Publish will allow me to, I think you can actually, this is new, I haven't tried it, you can like publish it into like the public Claude domains so other people can look at it like an artifact. you can actually copy and download the HTML file
you can share the data and the visuals because it's all in the HTML file. So you can do that, but it's probably not exactly what you were thinking.
Sarah (30:48)
for me, I would use this as a real start point and probably go off and do it in Tableau and maybe give it the image and bring it back in and say, hey, this is where I landed. And I find that's what I do a lot with Claude.
Darragh (31:03)
that's a great point because remember, could be great for wireframing. Typically when I wireframe stuff for work or clients, I'm using Figma or something like that. But if I've got the data already prepared, if I know the data we're going to use, I could just use something like this using MCP.
to connect to it and go, can you just like, give me a bunch of ideas around this data? This is what I'm thinking, test it out. do you want me to show you, I feed it an image of my dashboard from IronViz Did you wanna see that working?
Fiona (31:34)
Yeah, so
I would love to see what happens when you feed Claude a visualization and start your querying that way.
Darragh (31:45)
Awesome. So yeah, let's give this a go. So we've kind of explored this data set. Now I'm going to give it this image I prepared, a screenshot. So this is of my IronViz that uses this data set from nearly a year ago now. And you can't really see it here, but I might just bring it up screen
You can have a look, so I'll zoom in. So this is just the image of the Tableau public thing, how I used it, this data set to build up. You can see I use a radial, radial looks like a CD, that's where I was going from. And that's how it compares like positions on that. And then did a bunch of other analysis using these, radials and scatter plots and various other things, lots of analysis. So just to give you an idea, I've put this into Claude now.
I'm gonna ask it a fairly simple prompt. Like here is a visual dashboard that uses the hit machine data source. It's comprised of a number of visual elements. Can you do your best?
to replicate.
we'll just give it a whirl, see what happens. Hopefully, A, I don't run out of tokens and hopefully, nothing on my computer crashes, but we live and dream. So let's give it a go and see what happens. Now, I have tried this before and Claude kinda got somewhere there, but it kinda missed the mark a lot. So it'll be interesting to see what happens here. I wonder if it remembers from my previous conversations.
how it did it. So let's see how we go. So we got it. It's going to take a few minutes guys.
Fiona (33:20)
It's like, do you know what
this reminds me of? It's like the Wheel of Fortune. Like, what's it gonna land on?
Darragh (33:26)
Yeah, what's it gonna do?
But yeah, it's kind of interesting. I mean, while we're waiting for it to run, could you guys see how you might use it in your day to day?
Sarah (33:42)
me, it's really on what you said before around wireframing. Like I don't think I would be comfortable right now presenting this outside of probably Fi and I and taking it to clients, but it's a really good start point. And I think just the way that you can iterate so quickly and get it to look at it from very different angles to maybe what we would look at it like, and then quickly go, β actually,
Darragh (33:46)
Hmm.
Hmm.
You
Hmm.
Sarah (34:10)
that could go over here, that could go over there. it's got too much cognitive load and those colors don't work. But I think wireframing would be a really good use case for this for me.
Fiona (34:22)
When I think about how I would apply this in my day-to-day work, I'm thinking I want to feed it a bunch of design guidelines in the project space up front. having that in the instructions, I think that will help with some of the accessibility issues that might pop out. But really what I loved when you started the demo today was how it came up with some auto-generated insights.
And it's going to find those much more quickly than what I can possibly do by cycling through things. And I might be able to visualize things more effectively for now and helping to draw people through the story. But what it's doing is it's giving me the starter for 10. And then I can apply my own data visualization knowledge, which is the thing, you know, I've been doing for 20 odd years now.
Darragh (34:52)
you
Mm.
Mm-hmm.
Fiona (35:15)
and making sure that people can understand and interpret the information as it should be with the context that they require as well.
Darragh (35:24)
Hmm. That's very interesting. Yeah. It's an interesting space. but anyway, while we were talking, it has finished or at least it has kind of come out with something. Let's see the results. Um, yeah. What's the intern
Fiona (35:32)
The intern. see what the intern's done.
Sarah (35:38)
I like the way it's naturally just calmed the noise just by giving it your visualization, which is more best practice. It just feels a lot calmer. even before we get into what it's delivered, visually it's looking a lot cleaner now. So I think to your point, Fi if you could...
feed it all the guidelines, that would be a great, great start point as well.
Darragh (36:02)
Yeah, it's a good point. it's like, mean, my IronViz entry. I wouldn't say it's anywhere close to, best guidelines and data visualisation is very experimental. I saw it's got the, some of the colour scheme going as well. I use a purple and a pink and a blue. And then you can see it's figured that out from looking at the image,
We'll kind of have a look what's going here. As I start going down through the β visualizations, they're not really that close to what I had. So it's kind of not really nailed it, but...
if we had lots of time to discuss it, if you start giving it very specific examples and training it to go, β this is wrong and getting your prompts right, it does eventually get it closer. yeah, these don't quite work and these are definitely not in my original one. It's hallucinated them. This down the bottom though, this is interesting.
This is close to what I had in my original visualization. I'll just quickly bring it up if that's useful just to kind of show you that comparison. So it's, mean, it's kind of got the charts in the right ballpark. It's not the right data. So I'm not sure if you can see that. It's hard to show on the same screen, it's kind of
getting closer, the closer than I expected, let's just say. So yeah, what are your takes on that?
Fiona (37:18)
I love that it's got BANs at the top and it's also got a bit more context that's in there that it's obviously stolen from your visualization. So I think that works really well because that was one of the things that we obviously called out. As you scroll down, the natural data person in me is, okay, you've said that it's hallucinated. Has it really hallucinated or would we tag things like Dr. Dre and Snoop as,
Darragh (37:24)
Yeah.
Fiona (37:44)
rap or hip hop or however it's got all of those different categories in there. it just makes me even more curious about the data. These ranking changes, I know that there were more than that in the visualization. So it's a bit weird of what's going on. But it does give you a bit of an idea. Sarah?
Darragh (37:45)
Hmm.
Hmm. Yeah.
Hmm.
Hmm.
Sarah (38:06)
Yeah, again, it's a really good start point from, effectively dropping an image, getting access to a data source and giving it like a sentence. Like, if you continue to train it and, and it will evolve it's six months old, right, I'm interested to see where it goes.
Darragh (38:26)
So the large language models that we have, know, Claude, OpenAI, Gemini, whatever, they're trained on all this heaps of text data from, you know, millions and billions of documents. But the MCP protocol, the context protocol is essentially
The AI knows all that other stuff, but it doesn't necessarily know my specific data set. And to drag it back to the MCP conversation, what it's doing is using the context that already knows with this new context that we're bringing in and then joining it together. we're not, haven't pre-trained these models on my specific 1990s music data set, but I've added it into the brain of Claude.
And something you said there, Fi made me think of that because I was looking at the text up here and this text is doesn't make any sense in some ways. However, like I know I didn't write this, I'm pretty sure Claude has just used its other context from, its training process to build something out here. it's interesting how it works in that way. Like MCP is all about bringing that extra context into the Claude brain or the...
or the large language model brain.
Fiona (39:36)
Absolutely. we were chatting with, Chris Love a few months back now. And Chris was saying that he pulls all the information that he possibly can now into Notebook LM so that when he needs to recall something, he can just search query whatever he needs to and bring information together.
Darragh (39:52)
Mm.
Mm.
Fiona (39:58)
And it feels like this is sort in the same realms as well, but really giving us that access into data.
Darragh (40:03)
Yes.
Yeah, actually you just reminded me of something, Fi, I would recommend like if you or listeners or viewers are interested, there's a guy called Brian Julius on LinkedIn. He's actually a long-term like Power BI dev, but he's doing a lot of work in this kind of space in using MCP and BI and databases. And he does something similar to what you just mentioned that Chris was doing. He has this giant obsidian.
notebook like the obsidian if you don't know it's like an Evernote and he's got like basically anything he's ever written or thought it's just in there and he connects that via MCP to Claude and it's adding his entire dump of information that he's collected over the years into the AI system so he's doing something really similar to what you just described and yeah I think he's just doing some really interesting stuff in this space so
if you guys are both interested in it I do recommend reading some of what he's been putting out.
Sarah (41:01)
Yeah, we'll check him out. I'm just gonna ask you for a prediction here. How long before this type of interaction? Do you think becomes standard in business intelligence?
Darragh (41:13)
I mean, I don't want to panic anyone. think it's probably sooner than maybe we might think.
The fact that like Tableau have opened this up I think we're getting a steer from these big vendors in this space that they think it's something worth exploring. So I get the feeling it might be sooner than we may realize. How it evolves, I don't know. As I've talked about before, I still think, obviously there's going to be a place for
data professionals like us, but I get the feeling in a couple of years, we might not be doing maybe a lot of the grunt work and putting together visuals. We might be getting think Claude or another AI application to do a lot of that. Whereas we kind of tell it like what's best practice, what's guidelines,
doing that data storytelling part.
I get a feeling it's sooner than we realize, because there's a lot of interest in this area now. And we all know that the AI hype curve is higher than ever. for the record, I was fairly skeptical about AI. was just, yeah, it's cool, but it's just it helps me in ways that do make work easier. But it wasn't until I saw
doing things using MCP that I was like, actually maybe this is a bit of a turning point, particularly in business
Sarah (42:34)
Thank you. And I like the way you frame it as well. I optimistically think the same, that AI is here to help us do all the boring stuff that we don't want to do and allows us to do the other stuff. How that plays out, let's discuss this in six months to two years time, who knows?
Darragh (42:44)
Yeah.
Yeah,
no, exactly. I was pretty excited the first time I understood level of detail calculations in Tableau, but I'm probably pretty sick of them now. I could probably give Claude, give that to Claude to do.
Sarah (43:05)
yeah.
Fiona (43:07)
All right, Darragh that was amazing. Thank you so much. But before we wrap up, we want to do a quick lightning round with you. First thing that comes to mind, how often are you using AI in your day-to-day job?
Darragh (43:22)
β a lot. β Probably more than I imagined.
Fiona (43:26)
us too, so you're not alone.
Sarah (43:28)
What's the biggest limitation you've found with MCP so far?
Darragh (43:33)
it's a kind of weird way to answer this, but it's my own limitation in I don't know where exactly I can take it.
Fiona (43:40)
What is the most impressive query result that you've achieved?
Darragh (43:45)
I did get it to do a radial chart so in that example I showed before I did after a while of Persistent prompting I did get it to actually do a proper radial which is hard to do in Platforms like tableau. So I was quite impressed
Sarah (43:57)
In one word, describe the future of BI tools.
Darragh (44:02)
Possibly limited? That's two words, am I allowed to? If not, I'll go limited.
Fiona (44:05)
Okay.
Should data analysts be worried or excited?
Darragh (44:14)
excited for sure. Optimistic.
Sarah (44:18)
Next tool or platform you want to see MCP integrate with.
Darragh (44:25)
β I could be wrong here. I really like DuckDB. I'm not sure if you guys know that, but it might already be enabled. Someone's already done it, but that platform is pretty cool.
Fiona (44:38)
And last one, the percentage chance that this changes everything in the next two years.
Darragh (44:45)
I would say 90%. I won't go 100%, I'll go 90%.
Fiona (44:51)
Nice. Well, Darragh, thank you so much. It's been an incredible demonstration. We are so grateful for you showing us this glimpse into the future of data work. Any final thoughts on what we've just witnessed?
Darragh (45:06)
β no, thank you so much for your patience and yeah, it's been great chatting to both you Sarah and Fi. it's a very exciting space and I'm very, I feel very privileged that you've had me on to talk about it in still what is pretty early days in this space.
Sarah (45:22)
Now, if what we just witnessed excites you as much as it does us, make sure to check out Darragh's detailed blog. We'll link everything in the show notes below.
Fiona (45:32)
And you know the drill. If you enjoyed today's episode, please like and subscribe and most importantly, share this with your teams, your networks and anyone who's ready to see what the future of business intelligence looks like.
Sarah (45:47)
Until next time, stay curious, keep experimenting with your data, and thank you for joining us here on Undubbed, where we are unscripted, uncensored, and undeniably data.
