In-person obesity programs lack adequate digital support for blended care approaches. In addition, existing nutrition applications rarely integrate evidence-based personalization with clinical infrastructures. To address this gap, we present an early-stage prototype of an LLM-based conversational nutrition coach combining natural language interaction with a validated nutrient database and HL7 FHIR-based persistence. The system supports meal logging, personalized feedback, and goal tracking with interoperable storage. A formative usability study (n = 16) indicates high user acceptance, while revealing limitations in efficiency and consistency. These results demonstrate the feasibility of combining LLM-based interaction with interoperability standards for patient-centered nutrition therapy.