the problem
Models degrade quietly. Inputs shift, accuracy slips, and the first signal is often a complaint from a user.
what it does
driftline compares live traffic with a known baseline and tracks latency, input drift and output quality. When something moves, it says what changed and when it started.
how it’s built
A lightweight collector streams model events into a time-series store. Detection jobs run on a schedule, and alerts link straight to the affected window.
current status
In build. Collection and drift detection work today. Alert routing is next.
tech stack
- python
- node
- timescaledb
- grafana