About FinCanvas

Hi, my name is Marcos, I'm a software engineer and the creator of FinCanvas. I wanted to explore generative interfaces and came up with this little prototype.

We are now used to text-based interfaces when interacting with LLMs but I believe this might not be the right approach depending on how rich the data you have at hand is.

FinCanvas is still a chat-based interface but responses come back in widgets populated with the required data.

The project scope has been purposefully kept small, that's why data is completely mocked (aka invented) and widgets are very simple static cards that just display information. I believe though that this could easily be expanded into more dynamic workflows with real-time data.

If your're curious, the widget is generated in 2 phases:

1. LLM first determines which widget is the best fit for the message.
2. Then it extracts the data needed to render that widget, such as entities and metrics.

Thanks to this 2-tiered approach the system works fine with less-powerful LLM models. Right now I'm using chatgpt-04-mini with very satisfactory results.

Thank you for trying FinCanvas!