quietline: an open-weight model decides which nightly log warnings should wake me

Description and use

quietline uses an open-weight NVIDIA Nemotron 3.5 Lightning model to triage automation logs, sleeping through recoverable warnings and waking only for unresolved failures.

Problem

Scheduled jobs emit warning and error lines that often recover, making manual log review tedious and treating every WARN/ERROR as an alert overkill.

Outcome

In a synthetic test log, quietline correctly slept through recoverable 429 and cache miss incidents while flagging the failed build, demonstrating effective triage without unnecessary model calls.

Workflow

• 1. CLI reads new log lines and stores byte offset and hash of prefix.
• 2. Filters lines containing signal words (error, warning, timeout, 429).
• 3. Sends filtered lines with neighboring context to the open-weight model via NVIDIA NIM API.
• 4. Model returns "wake" or "sleep" decisions for each line.
• 5. CLI verifies decisions reference actual line numbers and prints original evidence beside any alert.
• 6. If no new signals, no model call is made.
• 7. State is persisted to allow incremental processing.

Evidence

quietline is a tiny Python CLI that reads only newly appended log lines and tries to wake me only for failures that still need attention.

Source

The model should sleep through the two recoveries and flag the failed build. On my captured run, it did exactly that.

Source

When the log has no new signals, it makes zero model calls.

Source