The Agenda

  1. Why everything I made this summer is terrible
  2. How AI has already made all of us obsolete
  3. My plan to ingratiate myself with our new robot overlords

Erik-generated district report

What goes into one of these reports?

  1. R code to crunch numbers (fine)
  2. R code to make crunched numbers available to Quarto docs (okay)
  3. R code to orchestrate the analysis and document generation (mildly irritating)
  4. The report template files in Quarto (rage-inducing)

What goes into one of these reports?

  1. R code to crunch numbers (fine)
  2. R code to make crunched numbers available to Quarto docs (okay)
  3. R code to orchestrate the analysis and document generation (mildly irritating)
  4. The report template files in Quarto (rage-inducing)

Making Report Templates with Quarto

  • Takes way too long

You’re going to get annoyed and use AI for most of the writing

  • The user is definitely going to change them

The user is going to use AI for most of the revisions

Hot Take: Technical writing isn’t a job for humans anymore.

The Robots are Here

Hand off tedious work to AI as soon as possible

  1. Crunch the numbers
  2. Store the results
  3. Pass the results to an AI agent along with a simple prompt:

Write a qmd file for an html document that can also be exported as docx. It should be a comprehensive report about chronic absenteeism in the district. Don’t editorialize, just report the results. It should be at least 1500 words long.

AI-Generated District Report


There are some drawbacks, but I’m brave enough to ignore them.

→
AI makes one document
(and boils the oceans)
→

The Real Challenge

  • Take the results of any analysis (not just absenteeism)
  • Store them in one of the usual formats (.csv, .rds, .dta)
  • Efficiently and reliably translate to LLM-readable format

Next Steps

Make the application that:

  • Parses data files into JSON for LLMs
  • Stores parsed data locally in a lightweight database
  • Generates custom prompts from filtered data
  • Handles APIs gracefully

Next Steps

Learn some 21st century skills:

  • Hardware needs for running LLMs locally (GPUs, RAM, etc.)
  • Secure local networks for offline, isolated deployment
  • Containerization and virtualization to isolate models from the sewer internet

Thank You