SRE / platform engineering → agent infrastructure. Shenzhen.
I build alerting systems, and the rails that let AI agents touch production without being trusted.
WebhookWise · self-hosted alert intelligence
Sits between your monitoring and your chat: dedup, AI triage, and a decision trace. "Why didn't I get paged?" gets an answer, not a guess.
It also caught itself being wrong. One week it filed 90% of alerts as high; an agentic investigator scored the same alerts and agreed with 26%. Measuring that dropped 59% of weekly volume out of high. The script still refuses to downgrade a rule the investigator called high more than a third of the time — that noise needs a more specific alert, not a mute.
That is the first thing in the README, ahead of the feature list. → How the calibration works (中文)
hookstack · agents on production, without being trusted
Three services; hookprobe runs one read-only agent per investigation. Read-only is constructed, not promised — scoped credentials, a bash guard, a disposable container.
The part I hadn't seen elsewhere is the input guard: an agent cannot write what steers its next run. One injected line in a runbook would otherwise outlive the run that read it. → Read-only by construction (English · 中文)
- Two alert systems, same production traffic, two days. 88.6% agreement across 35 comparable judgments. All four disagreements traced to one bug: recovery semantics lost across three system boundaries. → Data and method (中文)
- 13 PRs merged into easzlab/kubeasz (11k ★) — etcd backup paths, nodelocaldns upstream DNS, apiserver health checks behind ex-lb, harbor port probing.
📝 blog.wetalk.eu.org · 141 posts since 2018, mostly Chinese · 📫 imwl@live.com