Bridging Data Science and Generative AI

Episode Summary

The two halves of a data organization are converging. Domenic Ravita, VP of marketing at Plotly, describes generative AI teams starting to join the classical machine learning and data science teams they used to sit apart from, because both need the same data quality and the same preparation. What that convergence produces is not another reporting tool but a different kind of software: custom data applications, written in Python, that put machine learning, generative AI and a real interface in one place. His examples are a market data product sold to customers and a clinical trial operation that cut the time of the trial itself in half. His argument is that you cannot buy either one. There is no app off the shelf that does this, and the internal tools that promise to build it fall short, so the data science team that understands the problem has to build the application itself.

Key takeaways

  • His opening observation is organizational rather than technical, and it is the whole reason this conversation exists. The teams running generative AI experiments, which often sat as a separate group on another side of the company, are starting to join or work more closely with the traditional data science teams doing classical machine learning

  • The reason for the merger is unglamorous, and it is the part an executive can act on. Both groups rely on the same kinds of data quality and data preparation, so they want to share that infrastructure

  • He credits data scientists with their skepticism of generative AI rather than treating it as resistance. The skepticism is good, he says, and data scientists are in the job of reliable answers, or at least of being able to tell you how reliable an answer is; it can make them somewhat slower adopters for their mainline work in some cases, depending on the use case

  • Where generative AI is actually safe in a data science workflow is narrower than the marketing suggests. The use cases are usually the less risky ones, and can start in a developer’s own environment: data shaping, looking something up that would have meant going to documentation, searching your own GitHub repository

  • He draws the line at the output rather than at the tool. Generative AI is not relied on for a final answer to a customer or a final interpretation of an analysis, but it can be used in the workflow of creating one, and upstream of that in shaping the data

  • His history of analytics has a turning point, and it is not a product. The last decade of open data science is where the macro multi-decade trend of open source came fully home to roost, with most development happening in open source analytics code, mostly Python

  • He does not assert Python’s dominance; he points at the scoreboard. Every survey he names, the TIOBE survey and the Stack Overflow survey, points to Python as the leading language, growing beyond data science

  • The term he uses is bigger than the one the industry uses. Some people call this BI as code, but he says it is larger than BI itself: data analytics as code, which is a distinct change from a monolithic interface-first visual analytics tool

  • His one-line description of the product is the clearest thing in the episode. Plotly Dash is Tableau turned inside out and made accessible via Python

  • The convergence argument has a mechanical basis he states plainly. Everything being done in AI is done in Python, or mostly Python, so the AI ecosystem and the analytics as code ecosystem have come together

  • He extends the argument past the data team, and credits the idea to Apple. Everyone can code is mainly aimed at kindergarten through college, but he thinks any career will have some element of coding in it, because that is what lets you work with LLM-generated code

  • What the approach buys an executive is ownership, stated as three things rather than one. You can own the code, own the data, and own the AI capability you are realizing inside that function

  • The deployment reality behind that is specific and it is not a preference. Most of Plotly’s customers run it on-premise because they need it air-gapped or because it is a regulatory situation, so these are already sensitive data environments

  • The internal example is a pharma clinical trial application, and the business case is time rather than insight. A pharma customer built a custom data application for clinical data science that automates, orchestrates and analyzes the whole trial process, and the outcome he gives is cutting the time of the trial itself in half

  • He makes the central claim twice, and it is the one the episode turns on. There is not really an app off the shelf that can do this at scale effectively, so you have to build an app like this for yourself

  • The customer-facing example is a data product sold to customers rather than an internal tool. S&P Global’s Capital IQ, part of its market intelligence offering, is a Plotly Dash application that lets prospects and customers see what is in the data, explore it and run analysis on it, with generative AI summaries built in to help describe it

  • What the clinical trial application actually does is coordination, and he says the data science team supporting the trial is the one positioned to build it. It covers activation and the enrollment forecast, external benchmarks, site selection, patient selection, scenario planning and simulation of how the study is progressing, with metrics and KPIs across therapeutic areas

  • The outcome he gives for it is a decision about the company’s own boundary. Building that application allowed the company to basically insource the clinical research organization, which he says they spoke about at an open data science conference event

  • On agents he asks for a definition before he answers, which is unusual for a vendor. A lot of what is being called an agent today reminds him of heuristic-based automation, and he says there is nothing wrong with that, because AI-driven automation is where much of the productivity is

  • His verdict on autonomous agents is scoped to the stakes rather than to the technology. For a decision like the expensive clinical trial process he has just described, an agent that independently makes the decision is clearly not ready for those kinds of high-impact decisions

  • The reason he gives is a property of the systems, not a stage of maturity. Agents are non-deterministic, so where a deterministic result is needed the simpler and cheaper answer is an imperative tool or a statistical or classical machine learning score, rather than something that could be a hallucination

  • He does put agents in a sequence, which keeps the door open. AI-driven automation on open Python frameworks is the foundation on which a reliable agent-based workflow will be built in future, and in the meantime there is a lot of low-hanging fruit that might sound boring

  • The production number he cites is his recollection of a range, and it is why his company exists. Statistics or studies from a few different organizations put somewhere between 80 and 90 percent of data science projects as never reaching production, with different companies solving different blockers

  • His account of what a data application is becoming is a description of his own product’s direction. It is no longer only a web application with interactive visualization and read and write controls; it now incorporates generative AI and is starting to look like a compound AI system built from an open ecosystem of building blocks

  • He puts a name on why that has to be modular. A composable application built from building blocks is not something you can do with a monolith or a traditional interface-driven visual analytics product

  • His forecast for the enterprise is standardization rather than more experimentation, and he gives the reason. It feels a little bit like the Wild West in how companies are experimenting with generative AI, and the expectation at his company is that it becomes a little more operationalized and standardized, because it has to be for security reasons

  • The analogy that carries the second half of the episode is the one Christina singles out on air. The biggest general purpose LLMs are starting to become a kind of commodity, and he says a lot of the people implementing them now see them as akin to databases: you do not use one database type for every use case, you describe your workload and pick the one with the right price performance fit

  • The consequence he draws from it is a build-or-buy decision framed around competence. What you use depends on your business intent and where your core competencies are, which is what decides between something off the shelf from an external provider and running a model like a Llama 3.1 in-house

  • He is careful about which part of in-house is affordable, and the distinction matters. Running a model in-house is becoming cost-effective and retraining a large one would be expensive, but small models can be cost-effective to retrain, and in any case you can augment with RAG or with contextual documents in the context window

  • Asked for resources he gives two, and points past the home page to what is on it. plotly.com carries example applications in a gallery and user stories across the industries the company works in, and the second resource he gives is his own LinkedIn profile

  • He is explicit about where his company stops, which is rarer than it sounds. Plotly is fairly unopinionated about what you do upstream, the data engineering tool chain and the model development, including whatever MLOps technology you use to prepare the model

  • The problem he says the product exists to solve is a staffing gap, not a technology gap. Data science teams are not typically full-stack software developers, so building the application looks daunting, and the two alternatives are outsourcing it to another department or using in-house visual analytics tools that fall short

  • His closing advice to executives names the place to look rather than the thing to buy. Look inside the business for opportunities to create operational efficiencies and differentiating capabilities through custom data applications, for internal use and for customers, in the lab, the clinical trial and the field

  • The reason he gives for that advice is the same claim twice, and it is the note he ends the argument on. You are not going to find apps off the shelf to do these, and you are not going to find internal tools that help you build them

About Domenic Ravita

Domenic Ravita is VP of marketing at Plotly, the company behind Plotly visualization and Plotly Dash, the Python framework data science teams use to build interactive data applications. Years before, he worked as a field engineer with customers on real-time event data, merging it into actionable analytics in what he describes as monolithic, interface-driven visual analytics tools. His argument on this episode is that data science and generative AI have converged on the same Python ecosystem, and that the applications this makes possible, from a customer-facing market data product to an AI-driven clinical trial operation, are not available off the shelf and have to be built by the teams that understand the data. This is his first appearance on AI Realized.

 

In this episode

00:58 Welcome, and who Domenic Ravita is
01:28 The last time they spoke
01:40 The opening question: is this an exciting time for data scientists?
01:54 Why data scientists are attuned to what is happening
02:18 Generative AI teams joining the classical machine learning teams
02:38 The shared reason: the same data quality and the same preparation
03:06 Slower adoption for mainline work, and the Plotly community
03:28 Trusted results from probabilistic models
03:39 Starting with the less risky use cases
03:58 Data shaping, documentation, and searching your own repository
04:24 A thought partner, not the final answer
04:50 Upstream in the workflow, and where visualization comes in
05:05 How data and AI come together in code-based analytics
05:18 Field engineering, real-time event data, and monolithic tools
05:51 Tableau looked like magic, and then open data science happened
06:31 TIOBE and Stack Overflow, and Python as the leading language
06:58 BI as code, analytics as code, and Tableau turned inside out
07:27 Two ecosystems converging on Python
08:05 Everyone can code, and what it means for any career
08:40 A step function for operational decision-making
08:50 Owning the code, the data and the AI capability
08:59 On-premise, air-gapped, and regulated by default
09:29 Modular, open and extensible, and why that suits data assets
10:13 Examples from regulated environments
10:37 S&P Global, and what a market intelligence data product needs
11:08 Capital IQ as a Plotly Dash application
12:14 What a clinical trial costs in people, data and coordination
12:44 The custom application that cut trial time in half
13:46 No app off the shelf can do this at scale
14:24 Why the data science team is the one that has to build it
14:39 The question asked again, in different words
14:51 The same two examples, at greater length
15:17 Capital IQ, the data product sold directly to customers
16:31 Multiple datasets, and more than one way to manage the application
16:58 Pharmaceutical development, and customers like Amgen
17:24 Coordinating people, processes, data and analysis
18:13 Every step: enrollment forecasts, site selection, simulation
18:57 Earlier decisions, and insourcing the clinical research organization
19:20 The conference talk, and why he is bullish on this market
20:01 A case for agents?
20:04 What people mean by an agent, and heuristic-based automation
20:48 Not ready for high-impact decisions
21:16 Non-deterministic, and the cheaper deterministic answer
21:58 Low-hanging fruit, and the foundation agents will be built on
22:45 From pilots to production
23:12 80 to 90 percent of data science projects never get there
23:36 The application as a compound AI system
24:01 Composable applications, and what a monolith cannot do
24:48 Where AI is going for the enterprise
25:28 From the Wild West to operationalized and standardized
25:58 Local development, local AI, and LLMs as a commodity
26:29 Picking a model the way you pick a database
27:01 Specialized models, and models not built on language at all
27:45 What is cost-effective to run, retrain or augment in-house
28:11 Own your own intelligence, and keeping precious data inside
28:55 The same analogy applied to cloud and on-premise
29:23 Resources
29:31 plotly.com, the gallery and the user stories
30:07 Resources for choosing a model and designing the application
30:18 More sophisticated Python and gen AI components, and the focus on production
30:38 The last mile, and what comes together in the application itself
31:46 Data science teams are not full-stack developers
32:10 Outsourcing, BI tools, and why both fall short
32:49 What listeners should take away
33:27 Look for the operational opportunities inside the business
33:54 You will not find these apps, or the tools to build them
34:38 Close

In Domenic’s words

“Data scientists are in the job of reliable answers, or at least they can tell you how reliable an answer is.”

Domenic Ravita   (02:38)

“the way I like to describe Plotly Dash is if you took Tableau and turned it inside out and made it all accessible via Python”

Domenic Ravita   (06:58)

“they can cut the time in half of the trial itself”

Domenic Ravita   (12:44)

“there’s not really an app off the shelf that can do this at scale effectively”

Domenic Ravita   (13:46)

“it’s clearly not ready for those kinds of high impact decisions”

Domenic Ravita   (20:48)

“anywhere from 80 to 90% of data science projects don’t make it to production for various reasons”

Domenic Ravita   (23:12)

“It’s not one super LLM like a GPT-4 from OpenAI.”

Domenic Ravita   (26:29)

“You’re not gonna find internal tools that help you build that.”

Domenic Ravita   (33:54)


 

Resources

Domenic Ravita

Ideas and terms discussed

  • Analytics as code: His term for the shift, and he says it deliberately over the narrower BI as code, because what gets written in code is larger than business intelligence. It replaces the monolithic interface-first tool with something assembled from open source parts

  • Plotly Dash as Tableau turned inside out: The one-line description of the product. The point of the metaphor is access: everything the tool does is reachable from Python rather than from a fixed interface

  • Open data science: His name for the last decade in this field, which he describes as the macro multi-decade trend of open source coming fully home to roost in data science

  • Everyone can code: The idea he attributes to Apple, aimed mainly at kindergarten through college. His extension of it is that any career will have some element of coding in it, because that is what lets a person work with code an LLM generates

  • A custom data application: The thing the whole episode is about. Not a monolithic visual analytics tool and not a bought product: software built by a data science team for one problem, bringing a model, API calls, datasets and an interface people can act in together in one place

  • The last mile: Where he says his company operates, and where it stops. Upstream data engineering and model development are somebody else’s tools; the last mile is deploying, managing and securing the application and making it take actions

  • A compound AI system: What he says a data application is becoming, built from an open ecosystem of building blocks, which is why he also calls it composable. His argument is that a monolith cannot be assembled this way

  • Heuristic-based automation: His reading of much of what is called an agent today, offered without complaint: a lot of it reminds him of heuristic-based automation, and a great deal of productivity comes from automation, including AI-driven automation

  • Non-deterministic: The property that decides where he will and will not use an agent. When a deterministic answer is needed, an imperative tool or a classical machine learning score is simpler, cheaper and not a source of hallucination

  • LLMs as databases: The analogy Christina singles out on air. You do not use one database type for every workload; you describe the workload and pick for price and performance, and he expects model selection to work the same way

  • Local AI: What he says is coming, and what he talks about is cost. Running a model like a Llama 3.1 in-house is becoming cost-effective, small models can be retrained affordably, and RAG or contextual documents can cover the rest

Named on air

  • S&P Global and Capital IQ: His first customer example, named directly. Capital IQ is part of S&P Global’s market intelligence offering, and he says it is a Plotly Dash application, customer-facing, with generative AI summaries of datasets built in

  • Amgen: Named as an example of the customers Plotly has in pharmaceutical development. He does not attribute the clinical trial application he describes to any named company

  • UK Power Networks: Named once, at the very end, as a customer where the opportunity is operational decision-making in the field

  • Tableau: His own reference point, used twice: first for how magical monolithic visual analytics looked years earlier, then as the thing Plotly Dash turns inside out

  • TIOBE and Stack Overflow: The two surveys he cites for Python being the leading language and growing beyond data science

  • GPT-4, OpenAI and Llama 3.1: The models he names to make the database analogy concrete: not one super LLM, and a model you can bring in-house

  • GAI Insights and own your own intelligence: Raised by Christina Ellwood rather than by Domenic Ravita. GAI Insights is Paul Baier’s firm, and its model of own your own intelligence is the idea that a company should keep and build on its own data

Related AI Realized episodes and events

 

Frequently Asked Questions

 
 
 
 
 
 
 
 
 
 
 
Previous
Previous

When Attackers Have AI, Verify the Person Each Time

Next
Next

Treat the Agent as an Embedded Worker in the Ecosystem