Building an AI companion for a sensitive domain like mental health required more than a good chatbot — it demanded architectural safeguards against missed crisis signals, hallucinated advice, and slow responses. The system needed to reliably distinguish between everyday emotional support conversations and genuine self-harm or suicidal crisis indicators, respond to crises in a predictable and verified way, and track mood trends without letting inferred sentiment override the user's own self-reported state.

Designed a real-time LangGraph workflow with four specialized nodes. A risk classification engine built on OpenAI Structured Outputs evaluates every message for self-harm ideation or crisis indicators, tuned for recall over precision so potential crises are never overlooked, with any parsing failure or exception defaulting to high-risk as a fail-safe. When risk is detected, a template-driven crisis response system takes over immediately, guaranteeing consistent, hallucination-free delivery of verified resources like 988, Samaritans, and Crisis Text Line, with every crisis event logged for review. For non-crisis conversations, an empathetic AI companion agent offers brief, non-clinical guidance and evidence-based coping exercises such as breathing techniques, journaling prompts, and CBT-inspired reframing. A decoupled mood inference engine runs in the background, scoring sentiment from -1.0 to +1.0 independently of user-reported mood, keeping response times fast while preserving the user's self-evaluation as the primary source of truth. An administrative flag management dashboard adds human-in-the-loop review on top of the automated safety pipeline.

Designed a real-time LangGraph workflow with four specialized nodes. A risk classification engine built on OpenAI Structured Outputs evaluates every message for self-harm ideation or crisis indicators, tuned for recall over precision so potential crises are never overlooked, with any parsing failure or exception defaulting to high-risk as a fail-safe. When risk is detected, a template-driven crisis response system takes over immediately, guaranteeing consistent, hallucination-free delivery of verified resources like 988, Samaritans, and Crisis Text Line, with every crisis event logged for review. For non-crisis conversations, an empathetic AI companion agent offers brief, non-clinical guidance and evidence-based coping exercises such as breathing techniques, journaling prompts, and CBT-inspired reframing. A decoupled mood inference engine runs in the background, scoring sentiment from -1.0 to +1.0 independently of user-reported mood, keeping response times fast while preserving the user's self-evaluation as the primary source of truth. An administrative flag management dashboard adds human-in-the-loop review on top of the automated safety pipeline.

Building an AI companion for a sensitive domain like mental health required more than a good chatbot — it demanded architectural safeguards against missed crisis signals, hallucinated advice, and slow responses. The system needed to reliably distinguish between everyday emotional support conversations and genuine self-harm or suicidal crisis indicators, respond to crises in a predictable and verified way, and track mood trends without letting inferred sentiment override the user's own self-reported state.