GridOS：当电子表格遇上大语言模型，打造确定性 AI 财务建模引擎

GridOS is an innovative AI spreadsheet engine that pairs large language models (LLM) with a deterministic Python kernel. It addresses common issues in AI-generated spreadsheets—such as accidental data overwriting and formula errors—enabling users to edit spreadsheets via natural language conversations. Key highlights include reliable conflict resolution, multi-model support, and seamless Excel interoperability.

Spreadsheets are ubiquitous in business, but LLMs often disrupt existing structures when generating formulas or data—e.g., overwriting data validation rules or critical historical data. GridOS was developed to solve this problem by combining AI with a deterministic core to ensure data integrity.

GridOS uses a two-layer architecture:

1. Deterministic Kernel Layer: Built with Python, it includes modules for coordinate mapping, conflict detection, formula operations (SUM, MINUS, etc.), and A1 notation conversion. It ensures consistent outputs given the same inputs.
2. AI Agent Layer: A FastAPI app that streams table snapshots to LLMs, routes tasks to specialized agents, supports multiple providers (Gemini, Claude, Groq), and validates model outputs before application.

GridOS's conflict resolution engine prevents data damage:

1. Preview: Calculate potential changes.
2. Conflict Detection: Check for occupied/locked cells.
3. Auto-Shift: Move existing data to make space if conflicts exist.
4. Pre-Apply Checks: Avoid errors like division by zero.
5. User Confirmation: Let users approve/reject changes.

This workflow keeps users in control while allowing AI to suggest edits safely.

Features:

• Conversational editing (natural language to spreadsheet actions).
• Formula synthesis (e.g., "C3-D3" → =MINUS(C3,D3)).
• Persistent reasoning history (audit trail).
• User-defined macros and preset templates (DCF, budget, etc.).
• Excel import/export and cloud support (Supabase-based SaaS).

Use Cases: Financial modeling, budget planning, data analysis, education.

GridOS plans to add vector operations, external data connectors, provider-native structured outputs, and prompt caching. It represents a key direction in AI productivity tools—AI as a reliable assistant (not replacement) that maintains human control, making AI applicable in serious business scenarios.