Problem

The Bhote Koshi in Nepal rose about nine metres in half an hour under a clear sky. Every rain-triggered early-warning system was blind to it, because they all watch the wrong variable.

Approach

The detection lives in the database, not the app: a PostgreSQL statement trigger buckets gauge readings into five-minute windows, measures rate of rise against the last fifteen minutes, sums six hours of rainfall, and writes an alert in the same transaction as the insert. No cron, no worker, no polling. ClickHouse handles large-scale analytics, downstream village mapping tells a district officer who is at risk and how many minutes they have, and a LibreChat + MCP assistant turns raw alerts into language people can act on.

Results

In backtest the system warned 100 minutes before the surge reached Barhabise. 2nd Place at Better Days Hackathon (ClickHouse × LibreChat), San Francisco, Aug 2026.