Kurly Internalizes Multi-Building Address Refinement to Reduce External API Dependency and Achieve 99.2% Accuracy

Even after successfully handling single buildings, Kurly faced a problem where about 40% of traffic still depended on external vendor APIs, so it internalized its multi-building address refinement feature to solve this. Over roughly two months of development from January to March 2024, the company built logic that accurately matches various freely-typed forms of detailed addresses (dong, ho/unit number).

Accuracy Level Classification and Response Strategy

To increase the reliability of refinement results, OMS introduced the RefineAddressAccuracyLevel Enum, classifying results into 7 levels.

• EXACT: The ideal case where the exact dong/latitude-longitude is extracted from the same complex as the base address, accounting for 99.2% of all orders as of January 14, 2025.
• HIGH: Cases where the dong information is refined by tracing back from a neighboring complex different from the customer's entered base address. A search using the first 14 digits of the building management number and complex name tokens finds the uniquely matching building. This occurs at a daily average of 150 cases, and since the risk of delivery failure is high, a dedicated feature is needed.
• MEDIUM: Cases where a single representative building is returned in response due to errors in the Ministry of the Interior and Safety (MOIS) DB data or formatting differences (e.g., '208-dong' vs. 'two-hundred-eight-dong'). This occurs at a daily average of 300–400 cases, and continuous version patching is underway to reduce the possibility of misdelivery.
• BAD: Cases where detailed address information is missing due to customer input error, occurring at a daily average of about 100 cases. This is detected immediately upon payment completion, and the customer support team guides the customer to cancel and reorder.

Core Technology and Operational Tools

The core of resolving multi-building complexity lies in optimizing the matching logic with the MOIS Building DB. Preprocessing logic was applied to handle local government data quirks, such as mixed Korean/alphabetic numerals (e.g., '2o8') and notations including parentheses. In addition, the 'MOIS Building Management Page' within the OMS back office allows even non-developers to adjust building locations, look up nearby buildings, and correct data, maximizing operational efficiency.

Through this internalization, external vendor call costs were reduced to about 350,000 won per month, and Episode 6 will cover the process of proving this along with a development retrospective.