ROUTEAgents & automation
Jev-Oncall Alert Triage
Minglei@RaeX1101
Jev-Oncall uses Jev to determine if synthetic alerts are actionable, their severity, ownership, and if they are duplicates, routing them based on these probabilities or falling back to configured severity if Jev fails.
TakeCode turns those probabilities into routing decisions.
300 synthetic alerts. 7.6 seconds. About 1.3 cents in model cost.
I built jev-oncall with @typesafeai’s Jev to help answer a familiar on-call question: should this alert wake someone up?
One model call asks four questions:
• Is it actionable?
• How severe is it?
• Which team owns it?
• Is it a duplicate?
Code turns those probabilities into routing decisions. If Jev fails, routing falls back to configured severity.
In this run, 233 alerts reached Jev; rules handled 67 non-prod alerts. With 16 workers, median model-call latency was 418ms.
37% of judged alerts were flagged for human review. That’s the tradeoff I want to understand next: how much noise can I remove without missing something that matters?
github.com/mingleiw/jev-oncall
Related builds
