Services
AI systems fail in unfamiliar ways: a quiet model-quality drop, a vendor degradation, a cost runaway, a prompt-injection attempt. Standard SRE practice covers some of this; AI workloads need extensions.
How it works
Observability stack on Loki, Grafana, structlog by default.
SLOs that include quality, not just uptime.
Incident response practice — paging, runbooks, post-mortems.
Specific extensions for model-quality drops, vendor degradation, cost runaways, prompt-injection attempts.
Output
A working observability stack in your environment, with dashboards your team will actually open.
SLO definitions for the workloads that matter.
A paging and on-call rotation, set up to your cadence.
Runbooks for the most common AI-specific incident classes.
A post-mortem template and the first one filled in for a real incident (synthetic if needed for training).
Cost: TBC — engagement-based
