Product Engineer: Creating 5,000 Travel Products with Just a Few Clicks

MyRealTrip faced a bottleneck because tens of thousands of travel products coming in from partner suppliers had to be registered manually one by one. Creating a single product took dozens of minutes, and the maximum daily registration volume stayed around 20.

The solution to this problem was an automation system built by Kim Jerim, a product engineer on the T&A Experience team. The goals were automatic product registration, where simply entering a partner's product ID would convert it into a MyRealTrip product, and detail page decoration, which turned text-only information into something appealing to travelers.

The results were dramatic. The time to create a single product dropped to 30 seconds, and the daily registration capacity rose to over 5,000. It wasn't just faster—quality that used to vary widely by product was also standardized into a consistent layout and tone.

At the core were three roles played by AI.

• Translation and rewriting: Supplier data in English, Chinese, and Japanese was converted into Korean, not as a literal translation but rewritten into titles and copy designed to drive clicks.
• Automatic detail page HTML generation: The system analyzed partner responses to select an appropriate combination from 11 color themes and 2 layouts, adjusting sections based on the number of images.
• Automatic category classification: For standard categories, the allowed category tree was included in the system prompt for classification. For display categories, candidates were narrowed down using latitude/longitude, a region matcher API, and mapping tables before a final selection was made.

Quality control was handled through few-shot learning. Titles and detail pages from 50 to 60 actually well-selling products were provided as JSON examples, and the most similar cases—based on country, category, continent, and sales volume—were selected and inserted into the prompt to improve results.

Partner-specific characteristics were also reflected. The system was designed so that GYG generates titles suited to worldwide tours and tickets, while Japan Ticket generates detail pages centered on Japanese gourmet products, including menus and city names. Even with the same automation, the tone and structure vary depending on the partner's context.

Cost reduction was also an important task. Since creating a single product cost about 200 KRW, unchanging style references and rules were placed in the system prompt to maximize cache reuse, while only information that varies by product was sent as a separate request. Adoption of the batch API is also under consideration going forward.

The architecture was built using the Strategy Pattern. Adding a new partner only requires implementing a single strategy, and calls to the product automation API are queued in Redis before being consumed periodically by batches. Results processed in parallel by workers are delivered via Slack, including success/failure and cost reports.

It started as a simple script, but within a month and a half it grew into a platform. Phase 1 of mass registration is now complete, and future plans include expanding partner integrations, enhancing the manager dashboard, optimizing search exposure, reducing costs, and preparing more advanced detail pages that can extend even to merchant-listed products.