GS Retail and AWS Enhance GS SHOP Search and Recommendation Development Process with AI-DLC

GS Retail and AWS jointly applied the AI-DLC (AI-Driven Development Life Cycle) methodology to enhance the development process for the GS SHOP search and recommendation team. Previously, limitations in understanding natural language intent and inconsistencies in experimental approaches across teams made rapid validation and fair model comparison difficult.

Search Pipeline Optimization

Team A utilized Amazon Bedrock and OpenSearch Service to build a natural language search pipeline. They extracted attributes such as ingredients and efficacy using Claude Haiku 4.5 and Strands Agents, and implemented hybrid search combining BGE-M3 embeddings with BM25. Validation results based on 1.08 million product data points and 11,221 evaluation queries showed that while the p95 response time of 2.66 seconds met the target (3 seconds), the beauty category inclusion rate of 53% fell short of the target (80%). Analysis revealed that quality degradation caused by misalignment between sales metrics and relevance signals during the reranking stage was identified as a key improvement point.

Standardization of Recommendation Model Experiment Environment

Team B established a standard experiment environment enabling fair comparisons among data scientists. Based on 7.26M interaction logs and 210,000 user data points, they introduced the CandidateModel interface to create a common testbed. Over 40 model experiments were conducted, with data leakage verification ensuring reliability. Rather than selecting specific models, the focus was on determining the next development stage through clear decision label systems such as STOP, PARK, and KEEP_SOURCE.

Practical Value of AI-DLC

This project holds significance not for reducing development speed, but for standardizing the flow of problem definition, decomposition, and validation. Reproducibility and transparency were secured through the Inception, Construction, and Operation stages of AI-DLC, along with approval gates and audit logs. GS Retail plans to expand real-service application in the future through recommendation gateway integration and Reranking policy improvements.