Here Travel: Building an AI-Powered ERD Tool—Reducing Failed Calls from 61 to 3

The Ad Platform Development Team at Yeogi Eottae Company developed an ERD tool based on MCP (Model Context Protocol) to solve the issue of AI's non-deterministic outputs. Previously, they instructed AI to create ERDs directly, which resulted in inconsistent outputs and failures to adhere to team rules. To resolve this, they separated roles so that AI is responsible only for design and conversation, while the tool performs actual drawing, layout, and application of naming conventions.

Separation of Roles Between AI and Tool

By having the AI use the tool instead of drawing directly, consistent results were achieved. The AI defines tables and relationships through conversation with the user, while the tool receives this information and performs more sophisticated layouts than Mermaid or HTML. In particular, relationship lines are automatically placed in the empty spaces between tables, and the design allows Korean names, English names, types, and NULL status to be viewed at a glance. Features allowing humans to manually adjust layouts or lines were removed to ensure that the same ERD is generated regardless of who makes the request.

Tool Design Principles and Validation

Thirteen MCP tools were carefully configured to prevent AI confusion. Only one modification tool was included to ensure clear pathways, and the system was set to reject or ask for clarification on ambiguous requests to prevent AI from guessing. Additionally, English names and types are automatically assigned based on the tool's built-in glossary and domain rules to maintain consistency. In tests targeting the Claude Haiku, Sonnet, and Opus models, failed calls decreased from 61 to 3, and the number of tool calls decreased by 44%. This resulted in a reduction of ERD creation time from half a day to under 30 minutes.