[D] What is a simplistic but highly visual way to demonstrate a conversational language model?

Follow the full discussion on Reddit.
I am working on a research project where we train end-to-end task-oriented conversational agents (e.g., we solve the MultiWOZ benchmark) with novel techniques. We want to create a simple demonstration that can run on lower-end hardware and is highly visual to demonstrate the abilities of a TOD system to young audiences - e.g., high school students. At the same time, it should be complex enough to convey even to researchers that our models, when trained with data different than the toy problem, would provide substantial results. To understand me, I already have one idea, which however is not convincing enough: Imagine having a geometric shape on the screen, and you can move it to different points on the screen through the language model, e.g., you can have complex queries such as "Move the point twice as far from the edge as it is currently." We can autoplay the demo with different pre-set queries to move on a screen, and when someone wants to interact with it, they could replace the pre-set questions with their own. This is visual and interactive, as I want it. However, it is insufficient since it doesn't convey a strong message by being overly simplistic and only moving a shape on the screen. On the other hand, out-in-the-wild demos, such as ChatGPT, also don't work for us since, to be sensible, they require a lot of resources and are not visual enough by being text-only.