For Kakao Healthcare, delivering accessible, user-friendly, and smart healthcare services and managing a complex technical infrastructure are deeply intertwined. The company aims to accomplish the goal by developing new solutions that help make hyper-personalized recommendations drawn from large datasets and embedding that data into AI-powered platforms. 

Solutions such as the company's data platform (built with AlloyDB) to add value to existing hospital clinical data, and research portal HAYCORN to provide healthcare data services to consumers like pharmaceutical companies, were designed to help healthcare professionals make more informed decisions faster. Furthermore, Kakao Healthcare debuted its AI-based, mobile chronic disease management platform, PASTA (Personalized, Accessible, Supportive, Tech-enabled, and Affordable), to more directly support patients and users in their individual healthcare journeys.

The team's vision for PASTA has driven its growth from a diabetes care-focused app into "a super-smart health management platform that uses AI to understand people and help them make more informed choices," says Kyungdoc Kim, director, AI Transformation at Kakao Healthcare. At the heart of this evolution is the app's AI-powered chatbot. Whether users have health-related questions or simply need tech support, the team wanted to expand the chatbot's capabilities with a multi-layered agentic AI system. 

Working with Google Cloud Consulting, Kakao Healthcare was able to draw on foundational AI models on Model Garden to quickly iterate and develop a more accurate chatbot. By building on top of existing models, the company was able to deploy a more accurate agentic AI chatbot in just a few weeks.

To strengthen the guardrails, Kakao Healthcare built in protections against prompt breaking (attempting to induce the AI to generate restricted content or respond to non-compliant questions) and prompt hijacking (exploiting the chatbot to uncover security vulnerabilities in the system or manipulating app functions). 

These security precautions are part of several layers of easily customizable agents built into PASTA that run simultaneously: a health QA agent, an app QA agent, a website agent, and a security-focused agent. This ensures that no matter what question a user poses, there will be an expert agent ready to address their needs.

Standardized code designs also allow the team to adjust which datasets are drawn on by which agents. For example, simple questions may be handled by Gemini Flash to better manage cloud costs, while more complex ones are routed to Gemini Pro.