Blog
Working notes on automated operations reporting and the engineering behind the pipeline.
2026-10-01
Why four agents beat one prompt for weekly reporting
A single model call that “summarises the week” fails silently: you cannot tell whether a number was wrong, missing, or invented. Splitting collect, clean, analyze and report into separate agents makes every failure attributable.
2026-10-01
Anomaly detection that admits uncertainty
Z-scores and consecutive-drift checks catch the movements that matter. The hard part is honest language: a flagged drop has candidate causes, not a confirmed one.
2026-10-01
WoW vs YoY: which change actually deserves your Monday
Week-over-week tells you what moved. Year-over-year tells you whether it matters. Reporting both, with the base period stated, stops the seasonal false alarms.