The author argues that large language models are often oversized for specific tasks like log routing, where bounded decisions and structured outputs are needed rather than general-purpose capabilities. They describe building a log-triage system using Expanso and Jev that separates concerns: explicit rules handle routine events, while a specialized model answers focused questions about actionability and severity, with human-defined routing policies applying the results.
Expanso and Jev demonstrate a hybrid log-processing system combining deterministic rules with AI judgment. A severity-only baseline routes logs by level alone, while the pipeline adds contextual analysis—occurrence counts, time windows, and model scoring—to improve routing decisions. The system remains inspectable by showing baseline versus model-assisted decisions side-by-side.