📡 Idea Radar 每天 8:00 / 18:00 推送 · 18:00 生成精选
reddit · SideProject · 57 分 ← 返回列表

I built an incident response agent using Hindsight agent memory to stop on-call engineers from debugging recurring outages from scratch

Hey everyone! I built Déjà Vu ( dejavu-oncall ), an open-source incident response tool built to solve amnesia during production outages. When an outage hits at 2 AM, generic LLMs usually suggest generic fixes. The actual solution usually exists in an old postmortem that was ignored. Using Hindsight (Vectorize's agent memory engine), Déjà Vu retains past incident postmortems, symptoms, root causes, exact resolution steps, and what didn't work. When a new alert fires,

五维评分

付费意愿
50
需求真实度
50
个人可落地
78
竞争空间
35
新鲜度
80

AI 分析

痛点:值班工程师在凌晨处理告警时,无法快速召回历史事故的根因和修复步骤,只能从零排查,导致重复事故反复发生。

目标用户:SRE/on-call 工程师、平台工程团队、运维负责人,尤其是已有大量 postmortem 但检索困难的中型以上技术团队。

证据:When an outage hits at 2 AM, generic LLMs usually suggest generic fixes. The actual solution usually exists in an old postmortem that was ignored. ... stop on-call engineers from debugging recurring outages from scratch ... recalls similar past outages in ~140ms ... catching that 3 separate outages were caused by pgbouncer connection pool exhaustion after concurrency changes.

变现:开源自托管 + 云托管订阅;按坐席/月或按事件量收费;企业版支持 PagerDuty/Slack/ServiceNow 集成、权限审计、私有部署。

MVP:2周内做 Slack/PagerDuty 机器人:导入历史 postmortem 和告警,向量化后在新告警触发时返回相似事故、根因和已验证修复步骤,并允许反馈命中/未命中。

竞品:incident.io、Rootly、FireHydrant、PagerDuty AIOps、Datadog Incident Management、Blameless、Traversal

摘要:发帖人发布开源事故响应工具 Déjà Vu,用 Hindsight 记忆引擎存储历史 postmortem,并在新告警时约 140ms 召回相似事故和修复步骤,还能用 reflect 发现如 3 次 pgbouncer 连接池耗尽这类跨事故模式。项目使用 Next.js 15、TypeScript、Groq 等,但目前是开源展示,没有付费或收入数据,也未看到评论中的需求附和。事故记忆/RAG 是真实 B2B 痛点,但赛道已有 incident.io、Rootly、PagerDuty 等成熟竞品。