Technical Aspects of Deploying Large Language Models to Support Physical Activity and Well-being

This bachelor's thesis focuses on the design and implementation of an architecture for integrating Large Language Models (LLMs) with the mHealth platform HealthReact. The aim of the thesis was to create a system capable of generating personalized health recommendations based on user data collected by the HealthReact platform, emphasizing flexibility in LLM selection and efficient prompt management. The thesis first analyzes available LLMs and their providers (including OpenAI, Google, Groq, E-infra, OpenRouter) and compares relevant Python libraries for LLM orchestration (LangChain, LlamaIndex, CrewAI, Haystack). Furthermore, it compares platforms for prompt tracking and versioning, such as LangSmith and LangFuse. Based on this analysis, the LangChain library was chosen for practical implementation due to its modularity and multi-LLM support, and the LangFuse platform was selected for prompt management, versioning, and tracking. The practical part of the thesis describes the development of a Python application that communicates with the HealthReact platform's API to retrieve and aggregate user data (e.g., steps, heart rate, sleep) over defined time intervals. The implemented architecture allows for dynamic connection to various LLM providers. The solution also includes a graphical user interface created using Tkinter for managing and testing system and user prompts stored in LangFuse. The result of the thesis is a functional prototype system that demonstrates the integration of the HealthReact platform with various LLMs for generating personalized health recommendations and provides tools for effective prompt management and tracking within the context of mHealth applications.