CrisisTriageAI
Real-time multimodal triage platform that analyzes text, voice, and live phone streams to detect emotional distress and classify risk levels. Runs entirely locally with no external API dependencies.
Demonstrates production-grade ML engineering with privacy-first architecture and real-time streaming constraints.
Live Demo Preview
These clips show the system running locally in real time. Full interactive deployment is not publicly available due to privacy, safety, and cost constraints.
This demo video includes a brief graphic phrase related to self-harm. This platform was designed for emergency and high-risk scenarios, and the demonstration intentionally showcases how the system responds to such language.
Live transcription, prosody metrics, and risk classification updating in real time
Real-Time Processing
Session history, analytics aggregation, and pipeline observability
Dashboard and Session History
Screenshots
What I Built
- Designed modular FastAPI backend with WebSocket streaming for real-time audio and text processing
- Implemented local Whisper ASR transcription and Librosa-based prosody extraction (pitch, speech rate, pauses)
- Built custom DistilBERT-based neural classifier for risk-level classification with 73+ pytest tests
- Developed Next.js dashboard with live triage view, session history, and analytics aggregation
- Architected privacy-first system: all inference runs locally, ephemeral data handling, no cloud dependencies
Sub-second inference latency on local hardware; zero external API calls for core triage functionality
System Architecture
Loading diagram...
Overview
A privacy-first, research-grade system for real-time mental health triage. Combines speech recognition, prosody analysis, and neural text classification to produce actionable risk assessments.