Triage application logs before paging or deep analysis
Jev scores log events or batches for usefulness and urgency so noise is suppressed and only real issues reach an LLM or on-call.
Author's figure: Jev Logs README models a $1,000 monthly analysis bill falling to about $129 if 10% of logs still need deep analysis, and labels this a model, not a measured result.. Reported by the builder, not measured by us.
The use case
Small ops and platform teams either page on too much or pay a reasoning model to read health checks and cache hits. Jev scores each log record or collapsed batch for diagnostic value and urgency. Code maps the scores to suppress, watch, review, notify or page, and low-confidence results always go to review rather than acting. Every record stays in the archive. The on-call engineer still decides what to fix, and paging thresholds should be reviewed after each incident. Start in annotate-only mode so nothing is dropped.
Example workflow
- Export logs through an OpenTelemetry exporter or pull batches from Loki
- Collapse repeated lines into batches
- Ask Jev for diagnostic value, urgency and whether it is actionable
- Code maps scores to suppress, watch, review, notify or page
- On-call engineer reviews the review queue and tunes thresholds after incidents
Tools
- Jev (TypeSafe AI)
- OpenTelemetry
- Grafana Loki
- ntfy or PagerDuty
Risks and guardrails
- A suppressed batch can hide the start of an outage
- Thresholds drift as log formats change
- Modeled savings may not match real traffic
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- GitHub repo Jev Logs by reachjalil (OpenTelemetry log scoring): github.com/reachjalil/jevlogs
- GitHub repo jev-logtriage by jyatesdotdev (batch triage to suppress/notify/page): github.com/jyatesdotdev/jev-logtriage
- PagerDuty MCP server (official page): support.pagerduty.com/main/docs/pagerduty-mcp-server-integration-guide (checked 2026-11-09)