Transforming personalized shopping by unifying data and AI

Challenges of a fragmented data infrastructure

For KakaoStyle, adopting data-driven AI was not an option but a necessity. Improving the accuracy of personalized recommendations and search calls for the sophisticated use of user behavior analytics and product catalog data. Keeping pace with fast-changing fashion trends also required greater data freshness and faster model retraining. Through a range of initiatives, KakaoStyle introduced features such as "ZIGZAG Lens" and "similar-item comparison in the cart." These features help customers find products more easily and compare alternatives during checkout. KakaoStyle also expanded its delivery services and product selection to raise customer satisfaction.

However, the existing data infrastructure had limitations in reliably supporting these AI use cases. Data exploration, model training, and operations occurred in separate, disconnected environments, which hurt productivity. In turn, this undermined model performance and service quality. In particular, data such as new arrivals, stockouts, and price changes had to be reflected in near real time, but the legacy batch-centric pipeline could not keep data fresh. "The complex pipeline structure meant that quality issues often propagated to downstream data, and without an integrated monitoring system, it was difficult to detect anomalous data in advance," explained Jiyoon Park, Data Engineer at KakaoStyle.

There were also several issues on the infrastructure and governance side. Because permissions were fragmented across tools, it was hard to track data usage and lineage. This posed risks to both data reliability and security management. To address these limitations, KakaoStyle decided to adopt the Databricks Data + AI Platform, which unifies data management, analytics, and AI into a single, connected flow.

Consolidating data management and AI capabilities

With Databricks at the core, KakaoStyle consolidated its fragmented data infrastructure and built a system for running data analytics and AI in a single environment. The company migrated its data pipelines, metadata management, dashboards, and model serving to Databricks. As a result, it could reliably deliver business-critical data across the organization and eliminate decision-making gaps caused by data delays. This created an environment where teams across the organization could make fast, consistent decisions based on trustworthy data.

Streamlining analytics with Databricks Genie One

Adopting Databricks Genie One introduced a natural language–based analytics environment. This enabled business teams to explore and use data directly without going through the data team, significantly reducing repetitive analytics requests. At the same time, connecting everything from analysis to visualization in a single flow accelerated the creation and sharing of dashboards. This helped speed up core tasks such as customer behavior analysis and product operations optimization. Incident response also became faster, giving data engineers an environment where they could focus on core development rather than operational firefighting.

The way AI and ML models operated was also greatly simplified. Infrastructure management and deployment used to be time-consuming. With MLflow, the company can now deploy and operate models easily with simple configuration. Relieved from the burden of infrastructure management, ML engineers can focus on model performance and new use cases.

Enhancing collaboration and governance with Unity Catalog

Unity Catalog elevates data governance and team collaboration by centralizing data access permissions and management. This strengthens security and reliability and makes data flows easy to understand. At the same time, having data engineers, analysts, and data scientists collaborate in the same environment significantly improved efficiency across the organization.

"As we work to realize AI agent–based data democratization, we're building an environment that ensures both security and accessibility through Unity Catalog–based governance," explained Jiyoon Park.

Significant improvements in operational stability and productivity

Since adopting Databricks, KakaoStyle has significantly improved operational stability and day-to-day productivity—reducing incidents, accelerating analytics, and enabling faster self-service decision-making across teams.

• Batch failure alerts for the data engineering team fell by approximately 90% after migrating to Databricks serverless.

• Previously, resolving query syntax errors took anywhere from several minutes to tens of minutes, but with Lakeflow Jobs’ automated remediation feature, this was reduced to one to two minutes.

• For specific queries, the company achieved execution speeds roughly 2.6x faster than before.

• Dashboards that once took days to build can now be created quickly through prompts.

These stability gains translated into lower costs for maintaining and managing the previous clusters, as well as reduced incident-response staffing costs, lowering overall operating costs. The Genie One based self-service analytics environment also provides business teams with a foundation for making data-driven decisions without relying on analysts. This helps embed a data-centric culture across the entire company.

Building on these results, KakaoStyle continues to strengthen the core business capabilities central to fashion commerce, such as personalized recommendations, trend forecasting, and demand forecasting. In particular, the company is focused on building a system that can quickly translate near-real-time data processing and analytics into business strategy.

Looking ahead, KakaoStyle plans to advance its infrastructure—built on Databricks—so AI agents can operate safely with trustworthy data. "Databricks will serve as a core foundation as we continue to build an environment where anyone, regardless of role or technical skill, can use data together with AI to make decisions quickly," explained Jiyoon Park.