We tried to build an Excel error checker. To achieve that, we needed to actually understand the semantic structure of a spreadsheet first. So we built that, and it turned out to be the harder, more general problem.

The core issue: most real-world spreadsheets aren't relational tables. Merged cells, multi-level headers, multiple tables per sheet, totals mixed in with data. You can't just dump them to CSV and call it done. LLMs handle the easy cases but fall apart on complex workbooks at scale.

Our approach uses an agent-guided compilation pipeline that produces SQL-ready relational tables with full cell-level provenance. This demo visualizes what we do: https://storage.googleapis.com/deeptable-public/deeptable_an...

We have a handful of early customers but honestly don't know yet whether this is a real market or a niche problem. We're posting this to hear from people who've dealt with arbitrary spreadsheet ingestion. Whether you solved it, gave up, or are still living with the pain.

If you want to try it on your own files, email me (see my profile for my email) and I'll give you API access.