BigBinary built Neeti, an AI assistant for NeetoCal that answers plain-English scheduling questions by calling a fixed set of read-only Ruby tools rather than letting the model write raw SQL. The post covers the tool-based architecture (using RubyLLM and Gemini), a real bug where the model returned a wrong count because it used a paginated listing tool instead of a dedicated counting tool, and performance tuning by lowering Gemini's thinking effort from high to low — which cut response time by ~33% with no quality loss. Key design principles: the model never touches the database directly, all tools are read-only, every answer exposes a trace of which tools ran, and sampling vs. counting are strictly separated. The pattern is designed to generalize across all Neeto products.
