Recap & Sendoff

Safe & Maintainable AI Apps in Python

What we built today

	Section	Key idea
🧭	Foundations	Focus LLMs on strengths, engineer around the rest
	chatlas	One API for any LLM — chat, tools, streaming
🔧	Tools & agents	Real capabilities, with guardrails
📦	Structured data	Typed data from text, images, and PDFs
	shinychat	5 lines to a chatbot with tool displays
	querychat	Self-service analytics, safe by architecture
🧪	Evals & observability	Know if it works, and watch it in production

Callback to the “what we’ll build today” slide from the welcome deck. Walk through each row and remind them of the key takeaway. The arc: from “how to think about LLMs” to “a deployed analytics app” in 4 hours.

The core concepts

1. Focus LLMs on language → code
2. Give them tools and context
3. Keep humans in the loop

• Can get a long way with just context engineering + tools + transparency.
  • You probably don’t need RAG, multi-agent, fine-tuning, etc.

One more thing: beyond chat

Today’s examples were mostly chat based — but chatlas is really broader than that.

Chat is a great default UI for LLMs, but not the only one:

1. Invoice data extractor
2. Workout generator
3. Other examples

Where to go from here

📚 Docs

• chatlas
• shinychat
• querychat
• Shiny GenAI guide
• chatlas w/ raghilda

💬 Community

• GitHub Issues on each repo
• Posit Community Forum
• Discord / Shiny community

💪 Free tools

• Posit AI trial (no credit card)
• Deploy to Connect Cloud

Thank you! 🎉

bit.ly/scipy26

Carson Sievert · Posit