Hey HN! I built Banker.so, an AI agent that actually understands and writes Excel files.

I upload spreadsheets to ChatGPT/Claude all the time. CSV files work okay. But .xlsx files with formulas, multiple sheets, and pivot tables? They break constantly.

The problem: LLMs process data sequentially (1D), but spreadsheets are 2D structures. Cell B5 has meaning because it's in column B, row 5. Formulas reference ranges like A1:C10. Current AI tools miss these spatial relationships because they are designed for humans’ visual understanding.

What fails today:

- Complex Excel files with multiple linked sheets - Formulas and named ranges get corrupted or ignored - Pivot tables and conditional formatting disappear - Table structures with merged cells break completely

I implemented algorithms from the SpreadsheetLLM paper to parse spreadsheets properly. The agent understands:

- Table boundaries and headers - Cell relationships and formulas - Multiple sheets and references between them - Data types and formatting

You can find a video demo over here: https://youtu.be/hCTijOa3H9k?si=-GbX0ibmtXMVabAm

Also added OCR to convert PDFs and images directly into working spreadsheets.

Example use case: Upload your monthly financial report with 10 sheets, formulas, and pivot tables. Ask "which department exceeded budget?" Get the right answer because the agent understands your spreadsheet structure.

https://banker.so/

Try uploading a complex Excel file. The one that breaks other tools.

Give it a try and push it hard - I am sure there will be places I can improve it further. Looking forward to your feedback!