Kurly Improves Response Speed with AI Error Analyst 'Jarvis' Integrating Gemini and MCP

Kurly's delivery system team developed and is operating 'Jarvis', an AI assistant that automatically analyzes error logs using Gemini and MCP (Model Context Protocol). Previously, searching and interpreting logs took more than 5 minutes, which caused system stability issues, but after Jarvis was introduced, response speed improved significantly.

Initial Limitations and Prompt Engineering

Initially, the team attempted search-based automation, but accuracy was lacking for system-specific errors. To address this, they connected Gemini via LangChain and applied 5 prompt rules, including role definition and severity assessment, enabling cause analysis at the level of a senior developer.

Strengthening Context with MCP Adoption

The main reason team members ignored the early Jarvis's suggestions was the lack of actual code or business logic information. To solve this, they adopted MCP to connect the full error messages from Datadog with the related source code on GitHub. Provided with rich context, Jarvis began presenting accurate diagnoses and solutions in the real production environment, earning the trust of team members.

Future Plans and Results

Within one month of adoption, changes were observed in the Slack channel, with fewer simple log-check requests and faster sharing of specific diagnoses. Going forward, the goal is to use past Slack conversation data as training data to provide more field-relevant solutions.