English · 中文
SRE at Alibaba Cloud, based in Shenzhen. On GitHub since 2010.
By day I keep cloud-native infrastructure stable. The rest of my time goes into one long experiment: turning ops experience into tools and knowledge bases that people can read and agents can consume directly as RAG corpora.
Some of the numbers behind that: 4,750 production-ops documents in a single knowledge base, 432 support scenarios automated end to end, 94 real P0/P1 incident case studies, 959 runnable architecture demos.
Skimming? Start with these three: resolve-agent (agents doing real support work), klaw (one binary holding a whole K8s ops practice), kudig-database (the knowledge base everything else reads from).
- agent-up: feedback → attribution → release → rollback, a working loop for improving agents in production
- KUDIG / ResolveAgent: turning runtime traces of ops tools into structured, agent-readable knowledge
- open-geo: research notes on Generative Engine Optimization (GEO)
| Project | What it does |
|---|---|
| resolve-agent | Support agents on AgentScope that route by failure signature across FTA fault trees, wiki KBs, skills and code analysis. 432 scenarios live |
| agent-up | Agent config lifecycle: edit, approve, release, track, roll back. User feedback is auto-attributed into config-change suggestions |
| docsy (private) | Docs platform for the agent era. Markdown becomes machine-checkable research packs with 14 linter rules and a health score; holds up on a 21k-doc repo |
| sokr (private) | OKR and postmortem tool for ops teams. Incident reviews become graduable team skill assets. Zero deploy, runs in the browser |
| mcp-database | Self-hosted MCP server framework (mcp4coder). 37 servers, 200+ tools, one base class; adding a server means implementing two methods |
| skills-database | 455 job-competency skills across 27 roles. Sources traceable, browsable offline from a single file |
| Project | What it does |
|---|---|
| kudig-database | K8s + AI Infra ops knowledge base: 4,750 docs, 21 domains, 13 cloud vendors, plus a fault-tree engine with 16 top events and ~300 bottom events |
| ai-guru-database | Chinese full-stack AI knowledge base. 2,061 core docs (~19M characters), 700+ concept cards, from math foundations to agents in production |
| klaw | One binary with a console, diagnostics engine, ChatOps bot, alerting and SOS voice. 73 analyzers, sub-second event push |
| etcd-guardian | etcd backup/restore operator that covers the managed-cluster blind spot Velero can't reach. Works out of the box on Alibaba ACK, plugs into klaw |
| kubernetes-network-master | K8s network troubleshooting in a single binary (kubectl plugin). Hop-by-hop causal tracing from DNS down to CNI, 20+ subcommands, light eBPF inside |
| Project | What it does |
|---|---|
| open-tam | Multi-agent troubleshooting on AgentScope + MCP. Three-part reports (symptom, root cause, action); every step is a replayable trace, so the fault registry doubles as the eval set |
| kuba-database | 94 real P0/P1 incident reviews across 46 years and 47 companies, distilled into 12 root-cause classes, each case told from SRE, CRE, FDE and SA angles |
| cmd-database | A page per command for 1,100+ commands, plus 741 production notes. Generated from one YAML source, checked by a Go CLI |
| open-arch | Presales-architect copilot on AgentScope 2.0: clarify, design, diagram, review, deliver. Knowledge lives in auditable Markdown skill packs |
| open-demo | 959 runnable demos across 18 stacks for SAs, delivery and PMs. Each ships with an architecture diagram, a full tutorial and working artifacts |
| fde-scope | A field OS for forward-deployed engineers: 4 phases, 18 steps, 10 code-level gates, 7 of them from FAT, ISO 13849 and CE |
| repo-database | A curated arsenal for FDEs. 1,000+ repos scanned monthly, 115 structured dossiers kept, every entry CI-gated |
Everything else (100+ repos across 20+ orgs): allengaller.github.io/repos
- allengaller.github.io — home base
- Projects · Repos · Links
Always up for conversations about agents in production, Kubernetes failure analysis, or GEO. For anything project-specific, open an Issue on that repo; I read them all. Collaboration-wise, the best fits are Agent / RAG / MCP / CLI tools that have to survive production, and turning scattered docs into knowledge bases that keep paying off.
© Allen Galler · Shenzhen