The Data Itself Is the First Prompt in Vibe Analytics
Episode Summary
Vibe coding got a name and an audience of software engineers. Vibe analytics is the same idea pointed at anyone with a dataset, and Domenic Ravita, VP of marketing at Plotly, explains what changes. Plotly’s implementation drops the opening prompt entirely, because the dataset itself is the first prompt: the data goes in, the charts and controls come out, each with a plain English description. He is deliberate that this is not the chat paradigm, which the company tried the year before. The audience he names is not only analysts but the one billion knowledge workers who must eventually make sense of data. On messy data he is careful: there is hope AI reduces the data engineering burden to some degree, though he will not guess at the percentage. His guidance to executives: they need not build the AI systems themselves, and should look for the force multiplier rather than the thing that pulls a team into the weeds.
Key takeaways
Plotly’s open source has about 1.3 billion downloads. He gives the figure for the open source library and says most people who have been through a program in computer science or analytics will probably have heard of it and probably used it
Analytics as code is the movement AI is converging with. He describes analytics as code as what lets you go beyond drag and drop interfaces and express something more specific, and says the convergence of the AI trend with analytics as code is what creates the opportunity to make data analysis much easier, faster and much more broadly accessible to people with lesser skills or who are just learning
The audience is not only analysts: he puts it at the one billion knowledge workers. Asked whether someone who is not a data scientist could use the tool, he says 100 percent, that it is for you personally but also shareable, and that it is for any knowledge worker who at some point needs to review or analyze a set of data whether or not they consider themselves an analyst
His worked example is a security team hunting patterns it cannot name yet. He says Plotly’s own cybersecurity expert raised it: they are looking at logs all the time, they know the known patterns but do not yet know the unknowns, and dropping a log set in returns a narrative explaining what is in the data with interactive visualizations
The point is getting to a better question, not just to the answer. He says it can accelerate getting to those next better questions and get you faster to insights, and just before that, that because the bar to explore, shape and visualize the data is lowered to where it is done for you, you can focus on the next question, which often raises questions you had not thought to ask
Data scientists already have a name for the phase this speeds up: exploratory data analysis. He says AI definitely augments and accelerates that phase, and that what they are doing with Plotly Studio makes it even simpler and faster
Plotly Studio is deliberately not built on the chat paradigm, and they tried that first. He says the approach is not based around the chat paradigm, that they were approaching it that way the previous year, and that with a chat interface you just need a little bit more guidance about what the next thing is
The data itself is the first prompt. Asked whether he means you simply feed the data in and the tool organizes it into graphs from what it reads, he says that is exactly right, that you do not have to even give it that first prompt, and that the data itself is the first prompt
Vibe analytics is the analyst’s version of vibe coding, and the parent idea is Andrej Karpathy’s. He dates Karpathy’s vibe coding post to around late January or February 2025 and describes the idea as AI code generation being good enough to go with the feel of what you want and give an intent-based statement. He is explicit that vibe coding’s audience is software engineers, and turns to the analyst from there
He does not claim the term, and points at MIT Sloan for the research. At 13:54 he calls it a concept he thinks has been floating around in the last few months, and says that a researcher at MIT, Michael Schrage, who teaches at MIT Sloan, picked it up and dove deep on what it means to co-develop analyses with AI, published in MIT Sloan Management Review in July. He says Schrage studied three different scenarios and companies, looking at how far you could go, what tools you can use, whether the results are accurate and verifiable, and how much time is saved
He does not accept the AI data scientist framing when it is offered to him. Given the description that the tool becomes your partner in the analytic process and knows things about representing data that you do not, he says he thinks there are people exploring that kind of AI data scientist approach, and then describes what the Studio provides instead: a visual way along with natural language, much simpler than the drag and drop interface of today’s prevailing BI tools
Messy data is still where the work is. He says even data scientists, years into this, will say a large percentage of their work is dealing with the messy data, cleaning it, shaping it, aligning it and refining it, down to whether California is abbreviated with a capital CA or spelled out, so that you can bundle, aggregate and compare like to like
He refuses to put a number on how much AI removes the data engineering burden. He says there is hope that applying AI reduces the data engineering burden on analysts and data scientists to some degree, and that he will not venture a guess as to what that percentage is, while adding that it looks very good
The gains he does quantify are local ones: 20, 30 or 50 percent. He says they are finding lots of opportunities to apply AI in the micro, in the small, to make the overall process 20, 30 or 50 percent better, and that the general theme is compression
The user of the AI is the responsible party. He says we need to be aware of what the AI is doing, because in the end the user of the AI is the responsible party and you are the one sending it on the task
The privacy answer is a private LLM connection inside the customer’s own boundary. He says customers want to maintain their corporate data privacy, so an uploaded dataset stays within their secure boundary, calls to the model are private, and the model is not training on their data, which lets them adhere to their security and privacy constraints
Older models get cheap faster than frontier models do. He says everybody sees that the costs of LLMs are not coming down quite as quickly as expected, but that the costs of previous model versions are coming down quickly, so you do not have to always be at the cutting edge for every task, which is a way to hedge your bet on costs and still go with a trusted provider
He does not think large language models are designed to reason. He says that although reasoning is how they are talked about quite a lot, LLMs are great at mimicking the human mind’s ability to do associations, which he calls very sophisticated pattern matching across a multidimensional vector space, and he points to Gary Marcus and others as the people whose tests show it
He puts reasoning in neurosymbolic AI, and says he has not seen it in production. He describes the associative power of an LLM combined with the reasoning ability of formal logic as potentially very interesting, and says that from his reading it is still very much a research and development area in academia and he has not seen a successful production business application of it yet
His first piece of guidance to executives is that you do not have to build it all. He says a lot of companies are using the LLM as a building block to do something very specific to enable tasks around data analysis and data science, an approach they call AI native, and that what executives should evaluate is the next generation stack, the processes, and how to rearrange or adapt organizational structure to use AI
He calls the go-to-market engineer a little paradoxical. As a marketing leader himself he says he wants his marketers to leverage AI but does not want them building systems, does not want them at the low level of an AI agent SDK, and does not want them spending time debugging API calls
Look for the force multiplier, not the thing that pulls the team into the weeds. He says he wants his people working at a much higher level that allows AI to be force multipliers, that this is his own philosophy in marketing and go-to-market, and that he thinks it applies across other functional areas
Asked for one thing to remember, he comes back to the same idea. His closing line is to look for the force multiplication that AI can provide your team
About Domenic Ravita
Domenic Ravita is VP of marketing at Plotly, the data visualization company whose open source library he puts at about 1.3 billion downloads. He came to marketing from the technical side, working as a solutions engineer in the field on real-time event streaming and analytics technologies before moving from technical sales and solutions engineering into marketing over time. He had been exposed to Plotly years before joining, at a company doing visual analytics, and says what drew him was the chance to bring his solutions design experience to a go-to-market role. Plotly’s products include Plotly Dash, the framework the company built for data scientists who are comfortable with Python and machine learning but are not front-end or back-end developers, and Plotly Studio, the AI-native product this conversation is mostly about. This is his second appearance on AI Realized.
In this episode
| 00:42 | Welcome, and the guest introduction: VP of marketing at Plotly |
| 01:54 | About 1.3 billion downloads of the open source |
| 02:51 | The pain of doing analytics on funnel and cohort data |
| 04:18 | Analytics as code, and what the convergence opens up |
| 04:52 | Plotly Dash, the balance it strikes, and the data scientists it serves |
| 05:32 | AI magnifying that approach, and the opinionated framework |
| 06:03 | Could a non-expert use it, and is the result shareable |
| 06:41 | One hundred percent, and the one billion knowledge workers |
| 07:03 | The cybersecurity expert who does not yet know the unknowns |
| 07:54 | One-click deploy, and the bar being lowered |
| 08:24 | Getting to better questions faster |
| 09:19 | Exploratory data analysis, and what AI accelerates |
| 09:45 | Data apps, secured, deployed and scaled for an enterprise |
| 10:12 | Deliberately not the chat paradigm |
| 10:36 | Data ingested directly, and natural language specifications |
| 11:30 | Checking the description back: feed in the data, get the graphs |
| 12:09 | The data itself is the first prompt |
| 12:16 | Raising vibe analytics, and what inspired the term |
| 12:34 | Vibe coding, and Andrej Karpathy’s post |
| 13:26 | The AI taking care of the HTML, the CSS and the JavaScript |
| 13:54 | Vibe analytics as the analog, for the analyst rather than the engineer |
| 14:20 | Michael Schrage at MIT Sloan, and what he studied |
| 14:45 | Three scenarios, and whether the results can be verified |
| 15:33 | Whether vibe analytics means vibing with a data scientist |
| 16:13 | Not quite: what the Studio provides instead |
| 16:36 | A complex flow of clicks and drags, and what replaces it |
| 16:56 | Wrapping the vibe analytics step, and going further than it |
| 17:35 | Where this fits an enterprise AI program |
| 17:45 | Messy data, and whether California is CA or spelled out |
| 18:29 | Hundreds of datasets a night, and a deliberately wide quality spectrum |
| 19:09 | Reducing the data engineering burden, and the guess he will not make |
| 20:06 | Twenty, thirty, fifty percent better in the small |
| 20:33 | Piecemeal tools stitched together, and hiding the complexity |
| 21:42 | Applying AI securely, privately and ethically |
| 22:02 | The user of the AI is the responsible party |
| 22:27 | The private LLM, and the secure boundary |
| 23:01 | What to be clear about if you are building an AI system |
| 23:25 | Small model on the desktop, or large model in the cloud |
| 23:37 | A large language model, with an ensemble on the roadmap |
| 24:05 | Frontier costs and the falling cost of previous versions |
| 24:32 | Hedging your bet on cost with a trusted provider |
| 25:22 | Whether a neuro-symbolic model would suit the work |
| 25:30 | Large language models were not designed to reason |
| 26:53 | Where neurosymbolic AI fits, and the production application he has not seen |
| 27:27 | Guidance for executives leading AI initiatives |
| 27:44 | You do not have to build it all yourself |
| 28:12 | The next generation stack, the processes and the org chart |
| 29:01 | Why the go-to-market engineer is a little paradoxical |
| 29:30 | Force multipliers, not the weeds |
| 30:24 | Chris Parmer, Plotly’s co-founder and the creator of Dash |
| 31:03 | The Rise of Vibe Analytics event, and Brian Julius |
| 32:15 | The free trial, and the Analytics Vibethon hackathon |
| 32:49 | Look for the force multiplication AI can provide your team |
| 32:55 | Wrap-up |
In Domenic’s words
“You don’t have to even give it that first prompt. The data itself is the first prompt.”
— Domenic Ravita (12:09)
“Vibe analytics takes from that idea, which is what if it’s not for a software engineer? How does AI help a data analyst more quickly analyze data?”
— Domenic Ravita (13:54)
“It’s not based around the chat paradigm.”
— Domenic Ravita (10:12)
“There’s hope here, application of AI, to reduce the data engineering burden on analysts and data scientists to some degree. I, I won’t venture a guess as to what that percentage is.”
— Domenic Ravita (19:09)
“We’re finding lots of these opportunities to apply AI in the micro, in the small, so to speak, to make the overall process, like, 20, 30, 50% better.”
— Domenic Ravita (20:06)
“The user of the AI typ- is the responsible party. You’re the one really that’s sending it on this task.”
— Domenic Ravita (22:02)
“I don’t think large language models are designed to do any kind of reasoning, although that’s how they’re talked about quite a lot.”
— Domenic Ravita (25:30)
“As a marketing leader myself, I want my marketers to leverage AI, but I don’t want them building systems.”
— Domenic Ravita (29:01)
“Look for the force multiplier, not the thing that’s gonna pull you down into the weeds.”
— Domenic Ravita (29:30)
Resources
Domenic Ravita and Plotly
Domenic Ravita on LinkedIn: His profile. VP of marketing at Plotly, and a second-time guest on this show
Plotly: The company. He points listeners here at 32:15 for the free trial. It is one of two addresses he reads out on air, the other being the blog index at 30:24
Introducing Plotly Studio: The article he sends listeners to at 30:24, written by Plotly co-founder Chris Parmer and published 2 June 2025. He gives only the blog address on air, and this is the post that matches what he describes
The Rise of Vibe Analytics: The event he previews at 31:03, held online on 23 September 2025. Michael Schrage, Brian Julius and Chris Parmer all spoke, and so did he
The vibe analytics research
Vibe Analytics: Vibe Coding’s New Cousin Unlocks Insights: Michael Schrage in MIT Sloan Management Review, the piece he tells listeners to look up at 14:20. It sits behind a subscription. The free MIT Sloan explainer below covers the same work
What is vibe analytics?: MIT Sloan’s own short explainer on the same work, which reads without a subscription. It carries two of the examples: a telecom that surfaced more financially relevant insight in 90 minutes than it usually does in 90 days, and a cybersecurity firm
Ideas and terms discussed
Vibe analytics: The concept the page is named for, and one he is careful not to claim. Vibe coding, which he attributes to Andrej Karpathy in early 2025, lets a software engineer describe an intent and get a working first cut. Vibe analytics takes that idea and asks it of the analyst: how does AI help someone analyze data more quickly. He calls it a concept he thinks has been floating around for a few months, and points at MIT Sloan for the research rather than at himself
The data as the first prompt: Plotly Studio’s particular answer, and the sharpest line in the conversation. Rather than typing an intent, you supply the dataset and the tool generates the interactive visualizations and controls from what it reads, each with a plain English specification of what it does. He is explicit that this is not the chat paradigm, which the company had tried the year before, because a chat interface leaves people needing more guidance about the next step
Analytics as code: The movement he says AI is converging with: expressing an analysis in code rather than assembling it through drag and drop, which allows far more specificity in the interactivity, the aggregations and the controls. Plotly Dash is the company’s framework for it, aimed at data scientists comfortable with Python and machine learning but not with JavaScript or data management
Data apps: What Plotly calls the artifact the tool produces. The distinction he draws is that they do not stop at first insight: the application is secured, deployed and scaled for the needs of an enterprise, which is the part that separates an exploration from something a company can run
Exploratory data analysis: The name data scientists already use for the phase AI compresses, which he abbreviates as EDA on air. His claim is that AI definitely augments and accelerates this phase of the process, having said just before it that Plotly Studio, in the company’s own words, makes that even simpler and faster
AI native: His term for companies using the LLM as a building block to do something very specific, in this case for data analysis and data science, as opposed to organizations building the AI systems themselves
The private LLM connection: How Plotly answers the corporate privacy question. An uploaded dataset stays inside the customer’s secure boundary, calls to the model are private, and the model does not train on their data. His general advice for anyone building an AI system is to be clear about what data it uses, what is passed to the model, what gets shared and what security vectors that exposes
Neurosymbolic AI: Where he puts reasoning, as distinct from what LLMs do. He describes LLMs as extremely good at association, which he calls very sophisticated pattern matching across a multidimensional vector space, and says the combination of that associative power with the formal logic of a neurosymbolic system could be very interesting. He says that from his reading it is still very much a research and development area in academia, and that he has not seen a successful production business application of it yet, while noting that a few companies are working on it
The go-to-market engineer: The idea he pushes back on. He says companies are trying to turn marketers into engineers and that as a marketing leader he finds the notion a little paradoxical: he wants marketers leveraging AI at a high level of abstraction, not working an agent SDK or debugging API calls. He says his marketing ops team does have system development and operation in it, and that what he does not want is his marketers working at that level
Force multiplier: His closing prescription and his one takeaway. Look for the thing that multiplies what a team can do rather than the thing that pulls team members down into the weeds and turns them into systems engineers, and he says the principle applies across functional areas rather than only in marketing
Named on air
Chris Parmer: Plotly’s co-founder and the original creator of Plotly Dash, whom he credits at 30:24 with writing the Vibe Analytics article, and whose perspective on it he says is informed by 10 years of working on the problem, first as a research scientist building Plotly’s interactive data visualization and later with Dash. Parmer is the guest on episode 37
Andrej Karpathy: Credited with vibe coding, in a post he dates to around late January or February 2025. His summary of the idea is that AI code generation is good enough that you can go with the feel of what you want and give an intent-based statement, and that although Karpathy is a skilled AI engineer he is not a front-end developer, so the AI handled the HTML, the CSS and the JavaScript
Michael Schrage: The MIT researcher who teaches at MIT Sloan and, he says, dove deep on what vibe analytics means and what it is to co-develop analyses with AI, published in MIT Sloan Management Review in July
Brian Julius: Named at 31:03 as a top LinkedIn voice for data analysis and Power BI, a speaker at the September event and, he says, a prolific user of Plotly Studio
Gary Marcus: Named among the people who have done the work testing what large language models actually do, and whose tests he says show that LLMs mimic association rather than reason
Tabnine: A company he says Plotly previously worked with in the area of ensembles and small language models, named at 23:37 while he is explaining why Plotly runs a large model today
Ollama and vLLM: The two ways he names for running local model infrastructure now, vLLM for a more production-based scenario. He says many of Plotly’s users are not yet at that point
Cursor: Named as an example of the AI coding environments vibe coding is used with today, alongside AI code assistants. Vibe coding’s audience, he says, is software engineers
The Analytics Vibethon: Plotly’s hackathon, which he says started on 15 September and ran for a few weeks. The host tells him the name gives him away as a marketer
Tableau: Named by the host at 02:51, among the tools she had tried and found challenging for funnel and cohort work. His own contrast is with drag and drop BI interfaces in general, and he names no product
Frequently Asked Questions
-
Vibe analytics is the analyst’s version of vibe coding: using generative AI to explore and analyze a dataset by describing what you want, rather than by building the analysis yourself. Domenic Ravita, VP of marketing at Plotly, traces the parent idea to Andrej Karpathy, who he says posted about vibe coding around late January or February 2025, and he describes that idea as AI code generation being good enough that you can give an intent-based statement and get a workable first cut. Vibe coding’s audience is software engineers. Vibe analytics takes that idea and asks it of the analyst: how does AI help someone analyze data more quickly. He is careful not to claim the term, calling it a concept he thinks has been floating around for a few months, and points to Michael Schrage of MIT Sloan, who published on it in MIT Sloan Management Review in July.
-
Vibe analytics is aimed at the analyst rather than the software engineer, and in Plotly’s implementation it can start without a prompt at all. Domenic Ravita explains that vibe coding begins with an intent-based statement, the way Andrej Karpathy described it: you say you want a web app that does a particular thing, and the AI handles the HTML, the CSS and the JavaScript. Vibe analytics takes that idea and asks how AI helps a data analyst more quickly analyze data. Earlier in the conversation he had already gone a step further about his own product: you do not have to even give it that first prompt, because the data itself is the first prompt.
-
A knowledge worker with a dataset and no analytics training can use an AI analytics tool, and can share the result. Asked whether the output was only usable by the person who made it, Domenic Ravita answers 100 percent: it is for that person but also shareable, and puts the audience at the one billion knowledge workers in the world, on the grounds that at some point each of them needs to review or analyze a set of data whether or not they consider themselves an analyst. His worked example comes from Plotly’s own cybersecurity expert: security teams look at logs constantly, know the known patterns and do not yet know the unknowns, and dropping a log set in returns a narrative explaining what is in the data with interactive visualizations to explore it.
-
The data being the first prompt means an AI analytics tool reads the dataset and generates the charts and controls before you have asked it for anything. Domenic Ravita confirms it when the description is put back to him: you do not have to even give it that first prompt, because the data itself is the first prompt. The approach is deliberately not built around the chat paradigm, the approach Plotly had been taking the year before, because with a chat interface people need a little more guidance about what the next thing should be. Instead the data is ingested directly, interactive visualizations and components are created immediately, and each component carries a natural language specification, in plain English rather than any analytical language, describing what it does, whether it is a chart, a table or a filter control.
-
AI may reduce the data engineering burden without removing it, and Domenic Ravita will not put a number on it. He notes that even data scientists, years into this, say a large percentage of their work is dealing with messy data: cleaning it, shaping it, aligning it and refining it down to whether California is abbreviated with a capital CA or spelled out, so that like can be compared with like. He says there is hope that applying AI reduces the data engineering burden on analysts and data scientists to some degree, and that he will not venture a guess as to the percentage, though it looks very good. The gains he does put numbers to are smaller and more local: applying AI in the micro, in the small, to make the overall process 20, 30 or 50 percent better. He also points out that whether the generation completes at all is itself a measure of how messy the data is.
-
You route the analysis through a private model connection so the dataset never leaves your own secure boundary. Domenic Ravita says Plotly’s customers want to maintain their corporate data privacy, so an uploaded dataset stays inside that boundary, calls to the model are private, and the model is not training on their data, which lets them adhere to their security and privacy constraints. His broader point is about accountability rather than architecture: you need to be aware of what the AI is doing, because in the end the user of the AI is the responsible party and you are the one sending it on the task. If you are building an AI system, he says, be clear about what data it is using, what is being passed to the model, what gets shared and what security vectors could be exposed in the process.
-
Large language models were not designed to reason, on Domenic Ravita’s reading, and what they do instead is association. He argues LLMs are extremely good at mimicking the human mind’s ability to make associations, which he describes as very sophisticated pattern matching across a multidimensional vector space, and he points to Gary Marcus and others whose tests show it. His analogy is a memory you cannot quite place until a friend completes the statement. Reasoning, he says, is something different, and it is where neurosymbolic AI fits: an LLM’s associative power combined with the reasoning ability of formal logic. He is clear that from his reading this is still very much a research and development area in academia, and that he has not seen a successful production business application of it yet.
-
A company does not have to build its own AI analytics tooling, and Domenic Ravita’s advice is to spend the effort on the stack and the organization around it instead. His suggestion to executives is not to feel they have to build all of these AI systems themselves, because a lot of companies are already using the LLM as a building block to do something very specific for data analysis and data science, an approach he calls AI native. What he says they should evaluate instead is the next generation stack, the processes, and how to rearrange or adapt organizational structure to use AI, because, as he puts it, it is not just about the technology and you have to reorient your organization around AI. He is against pushing functional teams into engineering to get there: he calls the go-to-market engineer a little paradoxical, and says that as a marketing leader he wants his marketers leveraging AI rather than building systems, working an agent SDK or debugging API calls.
-
[00:42] Christina Ellwood: Welcome to AI Realized, the podcast for enterprise executives leading AI deployments. From tackling security, data, and operational challenges to navigating organizational transformation, AI deployment offers a unique opportunity to redesign our organizations from the inside out. I'm Christina Ellwood, your host for today's episode, and we're talking today with Domenic Ravita, the VP of marketing at Plotly. Welcome to the show, Domenic.
[01:10] Domenic Ravita: So glad to be here, Christina.
[01:12] Christina Ellwood: It's lovely to have a chance to catch up with you. I understand that you will-- you've-- they're doing a-- you just did a recent product announcement, so maybe we'll get a little snippet from you about that product announcement a little bit later in our call. Your background is in data, databases, data analytics, so data all the time. You're an expert in the data market, and you are a, an expert in data tools. And so it's understandable why Plotly's data visualization would appeal to you as a product, as a marketing leader. Tell me a little bit about your journey to arrive at Plotly and your work there.
[01:54] Domenic Ravita: Yeah. So Plotly is, I think most people who have been through a, a program in computer science or analytics will probably have heard of it. They've probably used it. We have about one point three billion downloads of our open source. And I'd been exposed to it years ag- years earlier at a previous company doing visual analytics. And that was when I was focused as a solutions engineer in the field working with real-time event streaming and analytics technologies. And I, I transitioned from technical sales and solutions engineering over time into marketing. And when I came across Plotly, I just thought this is an amazing company, an amazing opportunity to bridge what the, what my experience had been in solutions design and leverage that knowledge to help inform the go-to-market.
[02:51] Christina Ellwood: I, as a marketer as well, experienced the pain of trying to do analytics on funnel data and cohort data and so forth, and it's, l- let's just say it's quite challenging, and most marketing departments don't have a dedicated data scientist on their team. So I think there's a lot of promise in being able to empower executives with this type of tool to allow them to visualize data much more easily than you can with tools like Tableau or others that I have personally tried and used in the past. So I imagine that AI is having quite an impact on what is possible with data analytics. Can you help us understand how AI is being applied there?
[03:39] Domenic Ravita: Yeah, I think it really all begins with when you think about AI's training set, right? We've all read about how LLMs are trained on basically the Internet's data. And so that also means that it's trained on a lot of open source code. And increasingly, there, there are these sort of converging trends that are happening. There's the trend of LLMs being trained on a lot of this code, and then just within the data analytics space, because there's more data is accessible, more people need to analyze data in all sorts of different roles It needs to become more accessible. And AI, basically, the convergence of the AI trend with analytics as code, and I'll explain what that means, is basically creating this new opportunity to make data analysis much easier, faster, and much more broadly accessible with people with lesser skills or who are just learning. So analytics as code is basically the movement that's been happening over the last few years that enables you to go beyond just the drag and drop interfaces and express something more specific. And it could be the interaction, interactivity with analysis, with charts. It could be in the aggregations that you're doing and the controls. And so Plotly has been in this space for several years now with the Plotly Dash framework to provide that happy balance between ease of use, yet flexibility. And we've primarily been serving data scientists who are comfortable with Python and machine learning models and of course analytics, but they're not front-end developers, so they're not skilled in JavaScript, and they're not back-end developers who are skilled in data management. So Plotly Dash has solved that happy medium between the two as a framework that balances that. But AI basically just magnifies and it augments massively the ability to make that approach more accessible through the l- the combination of the LLM's world knowledge with an opinionated framework about what sort of task to do and what sort of app to create.
[06:03] Christina Ellwood: Does that mean that someone like myself, who is not a data scientist, nor am I a data expert, although I push data around, I'm not an expert in how to, to represent data visually. Could someone like myself use your AI visualization tool to take, uh, data I understand and am familiar with and have it be represented in a graph or a chart that is meaningful to me because I understand what, what questions I'm answering or need answered, and can then share it with other people? Or is it just for me personally?
[06:41] Domenic Ravita: 100%, it's for you personally, but also it's shareable. And absolutely it, it's for the situation of any knowledge worker who just has a set of data. You don't necessarily have to consider yourself an analyst, but out of the one billion knowledge workers in the world, at some point you're, you need to review or analyze a set of data. So this is for you and all of those people. And it's even for the situation where you're maybe not so familiar what's in the data, and I'll give you an example. We, like every company, is working with cybersecurity experts, and this is a use case that, that our cybersecurity expert raised, which is they're looking at all sorts of logs all the time. They're looking for new patterns, and it's always an arms race in security. And so often they are not quite sure. They know the known patterns, but they don't yet know the unknowns, and so they're looking for early signals for ways to detect that. And this is useful for them even in just taking a log set, dropping it in, and instantly getting a narrative e- that explains what's in the data with interactive visualizations to do that. And of course, yes, you can deploy with just with one click to po- cloud, so that's really super easy. So be- because this bar is lowered into understanding how to explore, shape the data, and visualize it to basically it's done for you, you now can focus on the next question. And often it raises questions you hadn't thought to ask about your data if you don't know the data set very well, like the cybersecurity example or what you're looking for. It can accelerate getting to those next better questions and get you faster to insights.
[08:32] Christina Ellwood: Yeah, we definitely saw that in product-led growth data where we were looking at cohorts and trying to understand who is truly stuck and who is just on vacation, right? What are these cohorts, these dead cohorts mean, or these seemingly dead cohorts, and what is the activity, anomalous activity behavior, and is it a positive or negative anomalous behavior? So we could divide people into different cohorts to understand where are they experiencing friction, or is it a signal that they need assistance or something like that. So it's not like your threat detection, uh, scenario where you're looking for the anomaly more than your... or the unknown pattern more than you're looking for the known pattern. Is that a good example of the kind of case that you could find u- unexpected insights?
[09:19] Domenic Ravita: Yeah, most definitely. I think in that early data exploration, AI is really effective at helping you, and we would say what we're doing with Plotly Studio makes that even simpler and faster. But data scientists would call this exploratory data analysis or EDA, and- AI definitely augments and accelerates this phase of the process. At Plotly, we go further than that. We take that initial application, which we call a data app, and we can secure it and deploy it, make sure that it scales, et cetera, for the, the needs of an enterprise. So we are applying AI in a way that accelerates that first insight with the data and enables a new kind of rapid iteration. And I guess importantly, we sh- I should say that, like, it's not based around the chat paradigm.
[10:17] Christina Ellwood: Mm.
[10:18] Domenic Ravita: Um, we previously were approaching it this way, like last year, and we're looking at the next iteration of what that capability would be, and just realized that with the chat interface, you just need a little bit more guidance about what's the next thing. And so our approach is to allow the data to be ingested directly, interactive visualizations and components to immediately be created, a- as well as natural language specifications. And so these specifications are, you know, they don't require you to learn any sort of analytical language. It's much simpler than that, and it's just English descriptions of what each component does, and it's by component that gets Whether it's a chart or a table or a filter control, those sorts of things. And then that's your first step to, to do the exploration after the first generation. And so we think this is a step change in itself in how AI is used in user interfaces and user interactions.
[11:30] Christina Ellwood: Interesting. So let me see if I am gathering the importance of what you're saying. So instead of me talking to a chat interface and loading my data into the chat or using the API to connect it and saying, "These are the things I wanna look at, give me tables, columns, and rows of this and that." Instead of doing that, I'm just feeding my data into this AI tool, and it's organizing it into the graphs for me, w- just based on what it's reading in the data, and then I can change it from there. Is that what you're saying?
[12:09] Domenic Ravita: That's exactly right. You don't have to even give it that first prompt. The data itself is the first prompt. Oh,
[12:16] Christina Ellwood: I see. I see. Okay. So this reminds me of something I wa- I've been reading lately, I've been, I wanted to bring up in our conversation, something that is called vibe analytics. We keep hearing about vibe coding. What is vibe analytics, and what inspired that term?
[12:34] Domenic Ravita: Yeah, I-- let's just start with what vibe coding is, right? This is an idea concept that Andrej Karpathy tweeted about earlier this year back, I think in Feb-- late January, February of 2025. And he bas-- the idea is basically that AI code generation is so good these days, and it's gotten better since then even, that he could just go with the feel of what he want. Just give an intent-based statement. "I would like to create a web app that does XYZ and is like an e-commerce store for pencil tops," whatever that is. And it would develop that first cut pretty quickly and pretty simply. And as a web app, he-- although he's obviously a skilled AI engineer, he's not a front-end developer. So the AI was taking care of the HTML, the CSS, the JavaScript, all that sort of thing. And he could focus on just, what do I want this interaction to be for my users, and what do I want the user interface to be? So that's the idea of vibe coding, and it's really intended, and as it's used today with AI code assistance and AI coding IDEs like Cursor, the audience are software engineers. And so it helps, it helps you accelerate your coding efforts. Vibe analytics takes from that idea, which is what if it's not for a software engineer? How does AI help a data analyst more quickly analyze data? So that's the analog is vibe analytics. And it, it's a ni- it's a concept I think that's been floating around in the last few months. But a researcher at MIT recently picked it up, Michael Schrage, who teaches at MIT Sloan, and he basically dove deep on what is vibe analytics and what does it mean to co-develop analyses with AI. And I'd encourage your listeners to, to, to look it up. It's in the MIT Sloan Management Review published back in July. And he basically did a study with three different scenarios and companies to take a look at how far could you go, what tools can you use, are the results accurate, can you verify the results, and how much time is saved? And but I think the important thing with the relation to vibe coding is like there is an acceleration there, and it does do a similar thing for analysts in that you can get to those better questions first and get to those first insights quicker. And so when we saw this, we're like, "Yeah, that sort of feels like the experience, the first experience we're aiming for in how we're trying to transform the data analytic process for our users."
[15:33] Christina Ellwood: I gathered that the vibe analytic part for someone who's not a data scientist is like you're vibing with a data scientist. Like the tool, the AI tool becomes your, your com- your partner in the analytic process, and it knows things about manipulating and representing data that you don't necessarily know how to do, but you know what you're looking for. And so there's a kind of interactivity about what the, what it's doing for the non-ana- analyst. Is that a good description?
[16:13] Domenic Ravita: I do think that there are, I think there are people exploring that kind of AI data scientist approach. What the Studio provides is a visual way along with natural language for the data, and that is a much simpler than today's prevailing BI tools which have a drag and drop interface. And so to recreate a chart or do an analysis you have to remember this complex flow of clicks and drags through a GUI. But this is a much-- this lowers the barrier and really compresses that cycle into just fast iterations with natural language and visualizations.
[16:56] Christina Ellwood: Yeah. I can really see there's a lot of applications independent of AI initiatives for sure, and this ability to take the first step in your sort of vibe analytic step and then to be able to go further than that to do analytics that are more s- and more sophisticated is really appealing. I'm wondering if there is a natural type of AI initiative that an enterprise might be working on that this data visualization capability is particularly well-suited to
[17:35] Domenic Ravita: That's a good question. I think that it's wherever you've got this readily available data. I think this is, and this is speaks to the big challenge. You may have data today that is a little bit messy, and I think that's the scenario in lots of situations. Even data scientists, even today, years into this, will say that there's a large percentage of their work that is the dealing with the messy data, cleaning up the data, shaping the data, aligning the data, refining it, making sure the syntax is like, is California abbreviated with capital CA or is it spelled out or all these sorts of things so that you can then bundle, aggregate, and compare like to like. So this approach with AI really helps in cleaning up the messy data and refining it.
[18:28] Christina Ellwood: Mm.
[18:29] Domenic Ravita: And it, in a large part, it's a good test of how messy or clean your data is. And so we experience this ourselves in our own QA testing cycle, 'cause we're testing hundreds and hundreds of data sets every night in our nightly builds, and we have a huge array and like spectrum of data quality across these sets purposely And we're finding that, and this is where AI is getting better and also our own application of it is getting better, in that we can be a lot more persi- permissive to allow messier and messier datasets come through. And so what that means is that there's hope here, application of AI, to reduce the data engineering burden on analysts and data scientists to some degree. I, I won't venture a guess as to what that percentage is, but it looks very good because in our own sets we can see is the, is Plotly's generation of the visualization and the insight, does it complete? Is it relevant? Has it compared the right variables? And that is also... That's a measure of the messiness of the data, because it can really interpolate a lot of things for you. It's combining the world knowledge of the LLM. That said, as part of the context that we're providing, we're deliberately doing and we're automating a lot of that data engineering and data cleanup for the user as part of the generation process. There, we're finding lots of these opportunities to apply AI in the micro, in the small, so to speak, to make the overall process, like, 20, 30, 50% better. And again, the general theme is this compression. And Michael Schrage talks about this in his studies of these three different groups, that it compresses that cycle. Now they're doing it with piecemeal tools stitched together. What we're-- What we've built is a system that leverages all of those learnings, hides that complexity, and just gives you a simple interface that anybody can use, and you don't have to know what's going on under the covers. But I think- People who work with data a lot will recognize as they experiment with this and throw messier and messier datasets at it, it just works, surprisingly.
[21:01] Christina Ellwood: Yeah, that's interesting. Of course, everybody is working with datasets that's using AI, and all- almost all the data that we have is dirty in some way or another. So that is a value that everyone working with an AI initiative can get from using th- this type of approach. Um, the visualization component, of course, is a different value proposition than that, and there certainly are plenty of use cases where the data visualization seems like to be, uh, there'd be a natural fit. But deploying AI isn't just a technical problem. It's also a cultural organization and, uh, organizational or even an ethical problem. What barriers do you think this-- Of those, which barriers do you think this helps to address?
[21:42] Domenic Ravita: Yeah, that's a great question. I think, first of all, everybody needs to be thinking about this because we need to be applying AI to our tasks and jobs, but we need to do it in a secure way that maintains privacy and, and in an ethical way. And we need to be aware of what the AI is doing 'cause on-- in the end, the user of the AI typ- is the responsible party. You're the one really that's sending it on this task. The-- what we find with our customers is in this, well, realm of having AI-powered data apps and data analysis, is they wanna maintain their data privacy, their corporate data, their privacy. So the way that we do this with Plotly is through their private LLM. And so this ensures that when they upload a dataset, it's all-- it's within their secure boundary. When there are calls to the LLM, those calls are private, and the LLM's not training on their data Uh, yet they can benefit from this result that it's within their secure boundary and the data is not being trained on, so they can adhere to their security and privacy constraints. And that's important, I think, for anybody working with AI. If you're building an AI system, you need to be really clear about what data is it using, what's being passed to the LLM, what gets shared, what are the security vectors that, that could be exposed in that process. So that-
[23:25] Christina Ellwood: Is Plotly Studio using a small language model that is part of the desktop download, or is it using a large language model that is in the cloud, or what's the model that you're using?
[23:37] Domenic Ravita: We're using a large language model. We are looking at, in the roadmap in the future as to what, how we might use an ensemble and small language models. And in fact, we previously worked with a company named Tabnine in that area that provides... I know we're one of the early innovators in AI coding LLMs. And that's something that is, is something we're looking at. At the moment, the advancements in the LLMs are coming so fast that the large language model approach is working fairly well. That said, everybody sees that the costs of LLMs aren't coming down quite as quickly as were expected. But the previous versions, models, depend- depending on what task we're doing, turns out to be fairly... Those costs are coming down quickly. So you don't have to always be at the cutting edge for every task. And so there's an opportunity there to basically hedge your bet on costs and still go with a trusted LLM provider. But as far as back to your question of a local LLM, we've talked to our users about that, and many of them aren't yet to the point where they're running local LLM infrastructure. Although there are some pretty good ways to do that now with Ollama and vLLM, actually, for a more production scen- based scenario. But yeah, as, as customers start to host their own LLM infrastructure, our own private LLM connection is ready to enable that.
[25:22] Christina Ellwood: Yeah, that makes a lot of sense. It would seem to me that a neuro-symbolic model would be appropriate for your work, too.
[25:30] Domenic Ravita: Yeah, I think that is really the answer to the reasoning thing that everybody's talking about. I don't think large language models are designed to do any kind of reasoning, although that's how they're talked about quite a lot. And it's-- a lot-- there's been a lot of work by people like Gary Marcus and others who are experts in this area, and their pretty objective test that you can see that this, that LLMs are great at basically mimicking the human mind's ability to do associations. That's, uh, it, it's amazing. Hey, we went to San Francisco last week and we went to the theater where? It can come up with plausible next word predictions of where that, what that was. And that's very much like when you haven't thought about a past memory in a while, and it feels somewhat familiar, but you can't quite put your finger on it, and then your friend completes the statement. That is very... LLMs do that association really well, and it's very sophisticated pattern matching across a multidimensional vector space. So that's great, and this is how lots of people are using that capability to build. But reasoning is something different than that, and that's w-where neurosymbolic AI fits. There are some companies working on that. I do think that's where the combination could be very interesting. The associative power of an LLM put together with the reasoning ability through formal logic of a neurosymbolic system. But from my reading of that, it's, that is still very much a research and development area in academia. I haven't-- There are a few companies out there that are doing it, but I haven't seen a successful production business application of it yet.
[27:27] Christina Ellwood: Perhaps that'll be something we're talking about the next time we have you on the show. Yeah. What is your guidance for executives who are leading AI initiatives as it relates to data visualization?
[27:44] Domenic Ravita: I think there's a lot of excitement of applying AI to analysis, data analysis, and data science. You don't have to build it all, I guess is the key thing. There are a lot of companies like Plotly and others who are using the LLM as a building block to do something very specific to enable these tasks around data analysis and data science. We're calling that AI native. And so I think sort of the evaluation matrix that executives should look at is about what's my next generation stack and what are my processes and how I rearrange my or adapt my organizational structure to utilize AI. All of these change management things you have to think about, because it's not just about the technology. You have to reorient your organization around AI So my suggestion is, first of all, don't feel like you have to build all of these AI systems yourself. And I think in every functional area of the business, we're seeing, like, there are companies with technologies that wanna turn marketers into engineers. There's this concept of a go-to-market engineer, and to me, that's a little paradoxical because as a marketing leader myself, I want my marketers to leverage AI, but I don't want them building systems. In my marketing ops team, yes, there is system development and operation, but I don't want it at a low level of using an AI agent SDK, and I don't want it them spending time debugging API calls. I want them working at a much higher level that allows AI to be force multipliers. And so that's my advice. That's my own philosophy within, in marketing and how I think about AI and go-to-market, and I think it applies across other functional areas. Look for the force multiplier, not the thing that's gonna pull you down into the weeds and make your team members have to become like AI system engineer or engineers building systems like that. It's... You should be looking at a higher level of abstraction.
[30:06] Christina Ellwood: Okay. That's really good advice, I think. And do you have some resources that you'd recommend to the executives who are listening to either learn more about that, like these higher level tools that are available, or to learn more about you and this new Plotly Studio product?
[30:24] Domenic Ravita: Yeah. I think if you're interested in what's happening at the forefront of how AI is changing data analysis, check out our article about Vibe Analytics. You'll find it at plotly.com/blog, and our co-founder Chris Parmer wrote this, and he's the original creator of Plotly Dash, the data app framework. And so his h- his perspective on vibe analysis and vibe coding is informed from 10 years of working and trying to solve this problem first as a research scientist, which was his initial creation of Plotly interactive data viz, and then later with Dash. So I think that's... You'll find that very interesting. And we're gonna be talking a lot more about this shift in the market with AI and data analysis at our event September 23rd, The Rise of Vibe Analytics. And in that event you'll hear from people like Brian Julius, who is a top LinkedIn voice for data analysis and Power BI. And he's doing really novel things around pushing the boundaries of AI Power data an- analytics, and he's a prolific user of Plotly Studio. So you'll learn about his experiences coming from his background, and we'll be introducing new technologies and products at that event.
[31:44] Christina Ellwood: Is that a virtual event, Domenic?
[31:46] Domenic Ravita: That's totally online. That's a virtual event, and we will have the Plotly community around the world with watch parties. And yeah, it's open to anyone, so you're welcome to join.
[31:57] Christina Ellwood: Great. Well, it sounds like we have some links to put into the show notes. We have the MIT Sloan paper. We have the link to your, uh, virtual event. Maybe we have a, a link about Plotly Studio and how to get a hold of that and use that. I assume it's available in some kind of form that people can try.
[32:15] Domenic Ravita: Absolutely. There is a free trial, so you can just go to plotly.com and you'll create a free trial there. And if you're interested in having a little fun and a little bit of competition, join our hackathon that's starting on September 15th. It runs for a few weeks, and we're calling it an Analytics Vibethon. So we're really- Right. ... this, yeah.
[32:38] Christina Ellwood: Very clever. I can tell you're a marketer. Thanks. All right, that's great. If there was just one thing you would like our listeners to remember from today, what would it be?
[32:49] Domenic Ravita: I would say look for the force multiplication that AI can provide your team.
[32:55] Christina Ellwood: Okay, great. Thank you, Domenic Ravita, the VP of marketing at Plotly, for joining the show today.
[33:01] Domenic Ravita: It's been a pleasure. Thanks for seeing me.