[pbronez]

My initial thought is "this can't possibly work."

We don't even have text to SQL working properly, and excel is so much messier than that.

What simplifying assumptions are you making about the spreadsheets people send you? How do you ensure correct results?

[alexliu518]

Very critical question. Excel is indeed more complex. Before analysis, Capalyze first preprocesses the data, which is crucial. We have designed a set of preprocessing algorithms that essentially focus on how to better identify the data suitable for analysis in Excel and clean and repair it. This process also leverages LLMs, as we found that LLMs perform quite well in recognizing table structures.