Practical AI for Data Science

Simon Couch - @simonpcouch

AI Core Team @ Posit

Median LinkedIn post (2026)

AI will change EVERYTHING about data science. Hear me out…👇

Click to expand

Median LinkedIn post (2026)

In AI discourse:

• The Data being Scienced can happily be sent straight to OpenAI’s servers
• Data science can be “one-shotted”
• “Close enough” is fine
• Tokens are cheap

AI will change EVERYTHING about data science. Hear me out…👇

Click to expand

LinkedIn vs. reality

In AI discourse:

• The Data being Scienced can happily be sent straight to OpenAI’s servers
• Data science can be “one-shotted”
• “Close enough” is fine
• Tokens are cheap

In reality:

• Data science happens on mostly sensitive / confidential data
• Data science is messy, subtle, and context-rich
• It’s (still) bad to be wrong
• Agents require infrastructure

I want to:

• Show you what’s possible via LLM APIs
• Help you imagine making it work in practice

What’s possible via APIs

Talk to LLMs via APIs

Web vs. API:

• Web: chatgpt.com, claude.ai, other “chat interfaces”
• API: Using Python, R, etc. with an API key

Talk to LLMs via APIs

1. Structured data

2. Tool calling

3. Coding agents

Talk to LLMs via APIs

library(ellmer)

chat <- chat_anthropic()

chat$chat("Who are you?")
#> Using model = "claude-sonnet-4-6".
#> 
#> I'm Claude, an AI assistant created by Anthropic. I'm here to help
#> with a wide variety of tasks like answering questions, helping with
#> analysis and research, creative writing, math and coding problems,
#> and having conversations. Is there something specific I can help
#> you with today?

Structured data

Structured data

# How would you extract name and age from this data?

prompts <- list(
  "I go by Alex. 42 years on this planet and counting.",
  "Pleased to meet you! I'm Jamal, age 27.",
  "They call me Li Wei. Nineteen years young.",
  "Fatima here. Just celebrated my 35th birthday last week.",
  "The name's Robert - 51 years old and proud of it.",
  "Kwame here - just hit the big 5-0 this year."
)

Structured data

chat <- chat_anthropic()
chat$chat("Extract the name and age from each sentence I give you")
chat$chat(prompts[[1]])
#> **Name:** Alex
#> **Age:** 42
chat$chat(prompts[[2]])
#> **Name:** Jamal
#> **Age:** 27
chat$chat(prompts[[3]])
#> **Name:** Li Wei
#> **Age:** 19

Structured data

chat$chat(prompts[[3]])
#> list(
#>   name = "Li Wei",
#>   age = 19
#> )

Structured data

type_person <- type_object(
  name = type_string(),
  age = type_number()
)

chat$chat_structured(prompts[[1]], type = type_person)
#> List of 2
#>  $ name: chr "Alex"
#>  $ age : int 42

Structured data

Tool calling

Tool calling

chat <- chat_anthropic()
chat$chat("What day is it today?")
#> I don't have access to real-time information, so I can't
#> tell you what day it is today. You can check your device's
#> calendar or clock for the current date.

Tool calling

today <- tool(
  function() Sys.Date(),
  name = "today",
  description = "Get today's date",
  arguments = list()
)
chat$register_tool(today)

Tool calling

chat$chat("What day is it today?")
#> ◯ [tool call] today()
#> ● #> "2026-05-14"
#> Today is May 14th, 2026. That's a Thursday.

Tool calling

Tool calling

Tool calling

Tool calling

Coding (agents)

Agent = LLM calling tools in a loop

Coding agents: querychat

Tool: write_sql_query

Coding agents: side::kick()

Tools: bash(), console() (like Copilot or Claude Code)

What’s possible via APIs

Making it work in practice

Raise your hand if…

• Your workplace has some approved, secure deployment of an LLM
• That LLM is frontier(ish): Claude Sonnet 4.5, GPT 5.2
• You’re able to access that deployment via a chatbot
• Someone on your team may have figured out how to connect to it via an API

OpenAI API compatible endpoints

In R, e.g. side::kick():

formals(chat_openai_compatible)
#> [snip]
#> 
#> $base_url
#> Sys.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1")
#> 
#> $api_key
#> openai_key()
#> 
#> [snip]
#> 
#> $api_headers
#> character()

Practical AI for data science