Hey George. The blog posts are mostly half-baked ideas, though a few are more baked than others.

Right now, I’ve got 4 projects online, set up so that if I push to GitHub, they get ‘deployed’. That makes it easy for me.

The four are:

  • This blog
  • A python package langgraph-codegen – a DSL of for experimenting with different AI agent architectures using langgraph. Mostly lets you just choose an architecture, then write functions, since that’s about the level I can deal with. It’s a library, I only use it in code, but I made a command line tool that shows what it can do: list examples show example graph code nodes conditions state
  • An architecture DSL web site – this is a web app that shows a few architectures with the DSL. You’ll get an error, because it’s not running all the time – $5/month only gets you so far… But it does come up in a few seconds, so when you get an error, try again. Here’s a screen shot: Architecture DSL
  • A really half-baked conversation evaluation web site – it won’t make much sense (bad UI), but its for doing a “what-does-this-human-know-about-this-other-human” evaluation. If you click on that “Bob-Shirley/Next” button a few times, then hit update, it shows what Bob knows about Shirley, and what Shirley knows about Bob. There’s a lot of activity in this area, so I wanted to compare them, based on which one does the best job at continuing the conversation. So you can evaluate things like “which of these approaches is best for getting necessary information from a human”. Here’s a screen shot: Theory of Mind Evaluation
  • Speaking of cheap, the other project I have and use, is a ‘wrapper’ for LLMs, so I can use cached answers instead of paying for yet another response to the same stupid prompt (I panicked one day when I spent almost a dollar).