mental-health-crisis-monitor

Mental Health Crisis Monitor: AI-powered early warning system that detects suicide risk in real-time chats, and streams APM traces to Datadog—all on GCP Cloud Run. Life is all, we got to LIVE!

Comment

Inspiration

  1. Rising mental health crises in India and world
  2. WHO report on suicide https://www.who.int/news-room/fact-sheets/detail/suicide
  3. IIT + IIM Graduates suicide cases https://www.consultancy.in/news/4168/young-mckinsey-consultant-commits-suicide-due-to-work-pressure.

What it does

This system monitors conversations for crisis signals in real-time, automatically escalates high-risk situations, and provides comprehensive operational insights—all while being completely free to deploy.

How we built it

FastAPI + Vertex ai and GCP Cloud Run. ddtrace-run in Dockerfile. Traffic generator simulates multiple sessions (5 HIGH risks). deploy.sh with DD_API_KEY.

Challenges I ran into

  1. Datadog Events: Not visible despite traffic (APM traces worked instead)
  2. Cloud Run Cold Starts: 2-3s latency on first request (pre-warmed with traffic generator)
  3. ML Accuracy vs Speed: Balanced 5 signals (sentiment+urgency) under 500ms response time

Accomplishments that we're proud of

Live demo: "suicide" → instant RED banner 5 HIGH risks from traffic generator Datadog APM: "High rate POST /chat" traces Free tier only. JSON configs exported.

What we learned

  1. Datadog Cloud Run: Direct integration via DD_API_KEY + ddtrace-run works (no Docker agent needed). APM traces > Events for submission proof. Traces > Events: POST /chat high rate visible instantly in APM. Events require explicit datadog_api.submit_event() calls + correct filters (service:mental-health-bot). Free Tier Reality: 1M spans/month sufficient for demo. crisis.score custom metrics auto-created from ddtrace.span() tags. JSON Export Pattern: monitors.json, dashboard.json, slos.json fully recreate setup → audit-proof submission.
  2. GCP Free Tier: Cloud Run scales 1000+ sessions automatically. Traffic generator pre-warms cold starts (2-3s → 200ms).
  3. Observability: crisis.score metrics + POST /chat traces visible immediately. JSON configs (monitors/dashboard) exportable for audits.

What's next for mental-health-crisis-monitor

Add voice analysis. family alerts on HIGH risk. Multi-language (Hindi/Tamil/Malayalam). Production SLOs

Built With

Share this project:

Updates