AI Mental Health & Mood Companion

A full-stack AI mental health and mood companion built with LangGraph, FastAPI, and Next.js. Every message runs through a real-time four-node workflow — risk classification, crisis response, companion conversation, and mood inference — designed around the principle that safety belongs in the architecture, not just the prompts.

Build a conversational AI for a sensitive domain that stays empathetic and useful in normal conversation, but never misses a genuine crisis signal and never improvises during one — while keeping mood analysis fast and independent of user self-reporting.

Architected a LangGraph pipeline with four specialized nodes: a Risk Classification Engine using OpenAI Structured Outputs tuned for recall over precision (any parse failure or exception fails safe to high-risk); a template-driven Crisis Response System that bypasses generation entirely to serve verified resources like 988, Samaritans, and Crisis Text Line, with automatic logging; an AI Companion Agent offering brief, non-clinical support and evidence-based coping exercises (breathing, journaling, CBT-inspired reframing); and a decoupled Mood Inference Engine that scores sentiment from -1.0 to +1.0 in the background without overriding the user's own mood reports. An admin dashboard supports human-in-the-loop review of flagged conversations.