Mansahay

The Challenge

1. Access Treatment Gap: Severe shortage of mental health professionals (0.75 psychiatrists per 100,000 people in India) makes timely intervention inaccessible, particularly during off-hours.
2. Privacy Paradox: Users are hesitant to transmit sensitive emotional data (suicidal ideation, trauma logs) to cloud-based server infrastructures.
3. Contextual Amnesia: Traditional conversational interfaces do not persist state or recall historical user metadata across sessions, preventing long-term therapeutic rapport.
4. Unimodal Blindness: Text-only interfaces fail to capture vocal and paralinguistic indicators of distress such as pitch variation, jitter, and energy.
5. Fragmented Care: Self-assessments, journaling, and professional consultations are siloed in separate applications.

Our Solution

1. Hybrid Architecture: Combines local inference (Ollama) on the client side for privacy-first operations with cloud-based reasoning (Gemini API) and RAG for clinical retrieval.
2. Episodic Memory (Digital Brain): Integrates ChromaDB vector database to store and retrieve user metadata and conversation history across sessions to resolve contextual amnesia.
3. Empath Engine: A dual-stream classification model fusing raw audio (Wav2Vec 2.0 XLSR-53) and semantic text embeddings (DistilRoBERTa) using cross-modal attention to detect distress markers.
4. Deterministic Guardrails (Privacy Sentinel): CPU-based regex and keyword interception running in <100ms to immediately halt generative output and trigger emergency response overlays during crisis detection.
5. Unified Ecosystem: Integrates self-assessments (PHQ-9, GAD-7), encrypted journaling, guided breathing, art therapy, focus games, and doctor booking.

The Result

1. Classification Efficacy: The Empath Engine achieved 68% overall accuracy on the IEMOCAP dataset (Anger F1: 0.77, Sadness F1: 0.73), successfully identifying masked depression.
2. Latency & Performance: Average response latency of 800ms for standard local queries and <100ms for crisis detection interception, supporting 50 concurrent streams.
3. RAG Precision: Document retrieval system attained a Mean Reciprocal Rank (MRR) of 0.87 using Vector + BM25 search with Cohere re-ranking.
4. Beta Testing Metrics: Evaluation with 95 participants across 20+ cities yielded an average usability score of 4.20/5.00, a Net Promoter Score (NPS) of +67, and 77.9% positive chat perception.