Access to mental health coaching is often limited by time, cost, and availability, leaving many without consistent daily support for stress management or personal growth. Advances in artificial intelligence (AI), particularly large language model (LLM)-based conversational agents, enable scalable and on-demand support. However, mental health is a safety-critical domain where failures can have serious consequences, and general-purpose LLMs are not designed to consistently enforce non-clinical boundaries, avoid harmful or misleading guidance, or respond appropriately to crisis cues. This work presents the design and implementation of a safety-first conversational support system that separates risk assessment from response generation, integrates structured memory for continuity, and enables human-in-the-loop evaluation and supervised alignment.