Building Custom GPTs to Solve Data-Related Problems with Adam Mico
It’s no longer news that AI has shaken the tech industry. Some people have fears for their jobs and future while others are steadily riding the AI wave. In this article, I chat with Adam Mico who is a Data Extraordinaire.
Adam has so far created five custom GPTs to solve data-related problems.
- Data Mockstar: Creates mock datasets.
- Viz Crique Pro: Gives feedback on data visualizations.
- tData Doctor: Provides Tableau advice for Server, Desktop, Prep, and Cloud.
- tBlueprint Navigator: Customer Success Guide based on Tableau Blueprint.
- Won’t You Be My Neighbor?: An empathetic and gentle chatbot
He is also the Principal, Data Visualization Enablement and Fluency at Moderna. He is a published author, a Dreamforce Golden Hoodie recipient, a 3x Tableau Visionary & 2x Michael W. Cristiani Community Leadership Award Winner.
Keep reading to learn more about his journey into data analysis, his creative process, ethics, and the future of work.
Joy: Hello, Adam! Thank you for taking the time to chat with me today. Do you mind sharing your background? How did you get started in data analysis?
Adam: My journey started with Excel and then I joined a project from 2003 to 2005 as a business analyst. I was introduced to Tableau in 2011 but didn’t receive my first data analyst role until 2017.
Being part of the #datafam community helped me upskill quite quickly, because I had the analytics background, but didn’t have an eye for design. So I improved my design quite a bit by working tirelessly on a number of initiatives and also collaborating with a number of people in the #datafam.
Joy: It seems analysis was in your background for the longest time, but your eye for design is more recent. You created VizCritique Pro, a GPT that provides feedback on data visualizations. Will you say this was a source of inspiration while creating VizCritique Pro?
Adam: So basically my inspiration to create VizCritique Pro was from a version I did for work. So I built a lot of custom GPTs at work and I figured out a way to make a public version of one that doesn’t use any of the instructions and stuff I developed at work.
A lot of people are afraid to ask other people for feedback because people are always so busy. They feel like they’re bugging them. So having a custom GPT that could provide professional feedback very quickly on a static visualization is super helpful for a lot of people, and I was just excited to be able to share a public version of that.
I think I’ve scored three Iron Viz in a row, and this is what I do at work on a day-to-day basis. So I kind of what makes a viz good and what makes a viz need improvement.
Joy: So, using VizCritique Pro is like getting feedback on a data viz from Adam himself! For people who are shy to ask for feedback, this is basically a dream come true.
Adam: Pretty much. Yeah, almost identical, except I wouldn’t score data vizzes out of 10. But the scoring part is fun, and I wanted to have a gamification applied to it as well.
The funniest part is scoring was the most complicated part of building that GPT. The reason is that depending on the criteria, it tends to score lower in some areas if you don’t instruct it to weigh things properly. So there was a lot of testing. I tested with over a hundred visualizations and then adjusted the scoring and the feedback to make sure the feedback was ethical and followed data visualization best practices, as well as a number of other things.
So the good thing about this GPT is that a jump score of 5 to 5.5 is not reflective of the score. Meaning that it’s a big deal going a half point. So each half-point on VizCritique Pro is a massive jump. So that’s another way that people need to think about it. So if you have a really good visualization and it comes out like a 7.1, and then you find a visualization that has a lot of challenges and it scores a 5.2, It doesn’t mean yours is just two points above it. It means there’s a big distinction between your visualizations.
But the scoring is kind of fun to get a quick gauge of how you’re doing, at least with the visualization you’re sharing. And it was a lot of fun to build it and weigh it and do whatever needed to be done to make it pretty equivalent to what I would score something.
Joy: So how does one get the best out of VizCritique Pro?
Adam: If you want really good custom feedback, you definitely want to make sure that you provide that context so it knows what to look at. So if you’re creating an infographic and there are words, you may want to mention that. If you’re creating a business dashboard, you may want to mention that they are brand colors.
Also, try to focus on the feedback from VizCritique Pro. It would not only score you but also advise on optimal chart types and better applications of the data.
Using VizCritique Pro is almost identical to receiving feedback from Adam.
Joy: I admire your creative process and the amount of effort you put into your work. I was about to ask questions about ethics and it seems you consider that before building anything. Your process shows that you care about doing stuff that’s valuable and impactful and not just doing it because you want to exist.
Adam: Oh, that means a lot, Joy. About ethics, I ensure my GPT works right. I also reduce “hallucinations” in the model and never steal anybody’s work. I don’t want to negatively impact somebody’s job. I only want these little system-specialized assistant geniuses to boost their work without friction.
Joy: Seeing that AI and data practitioners will be working hand in hand in the future. How do you see that future for us?
Adam: That’s a great question and it’s a scary question for a lot of people. The main thing that data analysts and data practitioners will need to focus on in the future is learning. Right now, GPTs aren’t strong enough to take over anyone’s job unless you’re kind of like a one-trick pony and, there are also ethical concerns about that (AI/GPTs) as well.
You really want to be able to understand more tools, more languages, and how to work with ChatGPT or LLMs. So you know how to make it work for you in the near future.
Joy: Interesting. So this is like the Industrial Revolution but with AI.
Adam: Yeah, the Gen AI revolution. We need to utilize it and leverage it to get the best outcomes. I think companies are going to be leaner in the future, meaning that they’re going to try to do more with less. And, you really need to use your creativity and all the tools in your arsenal.
Joy: We’ve had an amazing conversation so far. So if someone wants to build a custom GPT like you have, how should they go about it?
Adam: So initially you should be an expert in the field that you’re trying to build that GPT for. Anybody can build a GPT, but I don’t like GPTs that are built for fun and shared on a large scale. They could be dangerous, and ignore ethical issues.
Next is to create clear instructions. You can do this by asking ChatGPT to write a basic instruction script on what you intend to build. And then, you would work with that, look at that, and see what else needs to be addressed there. Other applications like GPT builder are really helpful in looking through what your instructions are and making sure that everything’s covered, including ethics in your instructions.
I strongly suggest against building from the exact scratch. Have a good idea. If you have knowledge of the documents, refer to them and your good idea. And then initially have ChatGPT draft one up, if you have access to ChatGPT Plus. If you don’t have access to ChatGPT Plus, you’re not going to be building custom GPTs using that platform.
So when you’re testing it, you want to make sure it adheres to your knowledge, that it really understands your instructions. If it doesn’t understand your instructions, you could work through them and make sure that it has a very clear understanding of what those instructions are.
ChatGPT recognizes Markdown code better. So, I would always suggest it to those writing instructions. It understands the Markdown code better because it’s structured and that’s just basically how GPTs communicate real language with people..
The other important one is if you want to make sure that it (your custom GPT) has a stronger opportunity to adhere to certain actions, make sure that you have all caps when it’s always or never, just to make sure that it’s clear to the GPT that you never want to perform an action or always want to perform an action. Also, include some sample responses to make sure that it’s coming off in the way that you intend.
And besides that, make sure that you include what persona you want, as far as a person responding, like you can make the persona, anything you want, but it has to fit what you’re doing for the GPT. So, for example, tData Doctor gives guidance on Tableau, so I wanted it to be a friendly, approachable, professor-type character.
The one thing I learned is that even if you have really good instructions for your GPT, it may adhere to those instructions, maybe 85% to 90% of the time, which is really frustrating, especially if you are trying to protect your work. So for example, I don’t want my instructions out there on the web and people copying and pasting and overloading ChatGPTs with the same GPTs I have.
So initially, if you’re just starting out with GPTs, I would suggest you build and share with your friends, only. And if it’s vetted out and, and it is everything you intended to be, and it’s for the better good, you should be able to release it at that point.
Joy: Thank you so much for your time, Adam. I appreciate it. And I love this conversation we’ve had.
Adam: I’m thankful that you reached out to me. It was great talking to you, Joy, and I can’t wait to catch up in the near future.
Stay data savvy,
Joy (Jhoie) Victor 💕
