English Version
I was a cargo van driver in China. 47 years old. No computer science background. No formal training in software engineering.
And yet, I built a production-grade multi-agent knowledge management system that manages 3,800+ verified knowledge assets — with executable verification protocols, confidence scoring, and a complete human-in-the-loop curation pipeline.
This is not a demo. This is not a hobby project. This is a real system, running daily, that I build and maintain with my own hands.
The system is called VRKG — Verifiable Retrieval Knowledge Graph. It was born from necessity: I needed AI to stop hallucinating on private knowledge. I couldn't afford to keep burning LLM tokens on repeated queries. So I built something that stores knowledge the way scientists store experimental data — with verification methods attached, confidence scores attached, and provenance preserved.
I learned about large language models, retrieval-augmented generation, knowledge graphs, multi-agent orchestration, and autonomous task execution — not from a textbook, but from building the actual thing.
Today, my system runs:
- Three-Body Architecture: Hermes (interaction body) + Evolver (exploration body) + OpenClaw (execution body) working in coordinated loops
- Verifiable Knowledge Assets: Gene units with
validate_commandfields that can be re-verified at any time - Capsule Procedures: Trigger-signal-driven executable knowledge bundles that fire when specific conditions match
- Confidence-Aware Retrieval: Every knowledge unit carries a confidence score; the system never serves high-stakes answers from low-confidence sources
- 3,800+ production knowledge units across AI operations, engineering decisions, and project history
I'm not a computer scientist. I'm not an academic. I'm a 47-year-old man who drove a cargo van for a living, and I taught myself enough to build something that computer science PhDs find interesting to talk about.
If that doesn't say something about what AI makes possible now — I don't know what does.
中文版
我是一个货拉拉司机。47 岁。没有任何计算机专业背景。
但我建了一套生产级的多智能体知识管理系统，管理着 3,800 多个经过验证的知识资产单元——每个单元都带着可执行的验证命令、置信度评分，以及完整的来源追溯链。
这不是演示，不是实验项目。这是一套每天都在真实运行的系统。
系统名叫 VRKG — Verifiable Retrieval Knowledge Graph，可验证检索知识图谱。它的诞生出于最朴素的需求：我需要 AI 在我的私有知识上不要产生幻觉；我负担不起每次查询都消耗 token 的成本。所以我建了一套机制，让知识像实验数据一样被存储——带着验证方法、带着置信度、带着来源。
我自学了大语言模型、检索增强生成、知识图谱、多智能体编排，以及自主任务执行——不是从课本，是从建这东西的过程里学的。
今天，我的系统运转着：
- 三体架构：Hermes（交互体）+ Evolver（探索体）+ OpenClaw（执行体），协调循环运作
- 可验证知识资产：Gene 单元自带
validate_command字段，可随时重新执行验证 - Capsule 过程胶囊：触发信号驱动型可执行知识束，条件匹配时自动执行
- 置信度感知检索：每个知识单元携带置信度评分，高风险答案绝不从低置信度来源中提取
- 3,800+ 生产知识单元，覆盖 AI 运维、工程决策与项目历史
我不是计算机科学家。我不是学者。我只是一个 47 岁的货车司机，为了生计奔波，然后自己学会了足够多的东西，建出了连计算机 PhD 都觉得值得聊一聊的东西。
如果这不能说明 AI 现在让什么成为可能——我不知道还有什么能说明。
- 🔵 redagent-vrkg — VRKG system documentation, Gene/Capsule format specs, SOP, and published sampling reports
- 🔵 EvoMap Node — Production AI agent node running Evolver v1.75.0, with ATP auto-buyer & auto-deliver enabled
- 🔵 Zenodo Publication — VRKG paper: "A Verification-Anchored Knowledge Asset System for AI Agents" (DOI: 10.5281/zenodo.19879022)
- Email: yesimagine@gmail.com
- GitHub: yesimagine-oss
- EvoMap Node:
node_b83d6e6008dce32f
"I don't teach you how to make money. I share how I try." "我不教你怎么赚钱，我愿意分享我怎么尝试赚钱。"