Clare & Me

Delivers AI-powered mental health conversations across three distinct communication channels (phone with voice-to-text transcription, WhatsApp messaging, SMS text) using a unified conversation state backend that maintains context across channel switches. The system routes incoming messages through a natural language understanding pipeline that classifies user intent (emotional support, coping strategy request, crisis signal detection) and generates contextually appropriate responses using a fine-tuned language model trained on mental health conversation patterns. Channel abstraction layer handles protocol-specific formatting (SMS character limits, WhatsApp media support, phone call duration constraints) while preserving conversation continuity.

Analyzes user messages using a multi-label text classification model trained on mental health conversation datasets to identify emotional states (anxiety, depression, loneliness, anger, grief, etc.) and situational context (work stress, relationship conflict, health anxiety). Based on detected emotional state, the system retrieves and recommends evidence-based coping strategies from a curated knowledge base (cognitive reframing techniques, grounding exercises, breathing patterns, behavioral activation suggestions) matched to the specific emotion and user context. Classification confidence scores determine whether to offer direct strategy recommendations or ask clarifying questions to improve accuracy.

Monitors incoming messages for linguistic markers of acute crisis (explicit suicidal ideation, self-harm intent, severe substance use, psychotic symptoms, acute trauma response) using a rule-based pattern matcher combined with a trained anomaly detection model that identifies unusual conversation patterns (rapid message escalation, emotional intensity spikes, topic shifts to harm). When crisis signals are detected above a confidence threshold, the system triggers an escalation workflow: generating a crisis-aware response, offering immediate resources (crisis hotline numbers, emergency contact options), and optionally routing to human review or emergency services depending on jurisdiction and user consent settings. The system maintains an audit log of all crisis detections for compliance and safety review.

Maintains conversation state across multiple messages and channel switches using a session store (Redis or DynamoDB) that persists user context, emotional history, and previous coping strategies discussed. The system implements a sliding context window that retains the last 10-20 messages (or ~2000 tokens) to provide coherent multi-turn conversation while managing memory constraints. When users switch channels (e.g., SMS to WhatsApp), the session lookup retrieves prior context and seamlessly continues the conversation. Session metadata includes user preferences (preferred coping strategies, communication style, crisis contact info), conversation tags (topics discussed, emotional themes), and timestamps for conversation analytics.

Implements a freemium model with tiered access using a usage metering system that tracks conversations per user (free tier: 5 conversations/month, paid: unlimited) and enforces rate limits via a token bucket algorithm. Free users receive full feature access (emotional support, coping strategies, crisis detection) but with conversation quotas; paid users unlock unlimited conversations and optional premium features (conversation export, progress tracking, therapist integration). The system uses phone number or WhatsApp ID as the user identifier for quota enforcement; quota resets occur on a monthly calendar basis. Upgrade prompts are triggered when users approach quota limits (e.g., 'You have 1 conversation remaining this month').

Generates automatic summaries of multi-turn conversations using extractive and abstractive summarization techniques (BART or T5 models fine-tuned on mental health conversations) to identify key emotional themes, discussed coping strategies, and user-reported outcomes. Summaries are stored in the session context and can be retrieved by users (in paid tier) to review conversation history without scrolling through full message logs. The system also tracks progress metrics over time (frequency of emotional states, coping strategy effectiveness ratings, user-reported mood trends) by aggregating summaries across multiple conversations, enabling users to visualize emotional patterns and treatment progress.

Tracks user interactions with recommended coping strategies (which strategies were tried, user feedback on effectiveness, follow-up emotional state) and uses this feedback to refine future recommendations via collaborative filtering and contextual bandit algorithms. The system maintains a user-strategy interaction matrix where each user has implicit and explicit ratings for strategies (tried and reported helpful, tried but unhelpful, not tried). When recommending strategies, the system balances exploitation (recommending strategies with high historical effectiveness for this user) with exploration (suggesting new strategies to expand the user's toolkit). Recommendations are contextualized by emotional state, time of day, and previous conversation patterns.

Generates contextually appropriate, empathetic responses to user messages using a large language model (likely GPT-3.5 or similar) fine-tuned on mental health conversation datasets to adopt a supportive tone, validate emotions, and avoid harmful language. The generation pipeline includes prompt engineering (system prompt specifying role as supportive AI, constraints on medical advice), response filtering to remove harmful content (suicide methods, medication dosing, diagnostic claims), and tone adjustment to match user communication style (formal vs casual, verbose vs concise). The system uses temperature and top-p sampling to balance response diversity (avoiding repetitive canned responses) with consistency (ensuring responses stay on-topic and emotionally appropriate).

Learns user communication preferences (verbosity, formality, emotional tone, language) from conversation history and adapts response generation to match. The system tracks metadata about user messages (average message length, use of formal language, emotional intensity, emoji usage) and adjusts generation parameters (response length, vocabulary complexity, tone) to mirror the user's style. For example, a user who sends short, casual messages receives concise, informal responses; a user who writes long, detailed messages receives more comprehensive responses. Preference learning is implicit (inferred from conversation patterns) rather than explicit (requiring user configuration), reducing friction.