Open-Source AI Spreadsheet That Doesn't Hallucinate Math
The author built an open-source AI spreadsheet called GridOS that separates reasoning from strict arithmetic to address the problem of LLMs hallucinating math and breaking formulas.
Why it matters
This open-source project provides a solution to the problem of LLMs hallucinating math in financial modeling and complex spreadsheet applications.
Key Points
- 1GridOS acts as a deterministic spreadsheet kernel that separates the reasoning (handled by an LLM) from the execution (handled by a strict Python AST backend)
- 2The AI is physically blocked from writing directly to the cells, and every response goes through a preview step to check for collisions, locks, and formula integrity
- 3The core is MIT licensed and includes a drop-in plugin system to register custom formulas locally
- 4The author is stress-testing the AST parser to find and fix edge cases
Details
The author noticed that when using large language models (LLMs) to build financial models or complex spreadsheets, the LLMs often hallucinate math, create circular references, or silently break formulas. To address this issue, the author built an open-source AI spreadsheet called GridOS that separates the reasoning (handled by the LLM) from the strict arithmetic (handled by a deterministic Python kernel). The architecture includes a preview step that checks for collisions, locks, and formula integrity before executing the math. The core of GridOS is MIT licensed and includes a plugin system to add custom formulas. The author is currently stress-testing the AST parser to find and fix any edge cases that may break the engine.
No comments yet
Be the first to comment