Why I Built This Website: An HVAC Engineer's Experiment with AI Agents为什么我建了这个网站：一个暖通工程师的 AI 协作实验
When ChatGPT first arrived, followed by the rapid rollout of GPT-4, Claude, and Gemini, models became dramatically more capable with each release. As an engineer accustomed to working with the physical world, I was genuinely curious: what is the actual envelope of this tool—and how far can it go as a real working partner?
Over the past 15 years, my HVAC career has spanned a wide spectrum of built environments—from industrial plants and commercial spaces to hotels, offices, and residential buildings. I have also had my share of data center projects over the years: from precision cooling (CRAC/CRAH) in telecom equipment rooms and traditional internet data centers, to two-phase CO₂ cooling for high-density racks (a decade ago, 12 kW per rack was already considered extreme density). Today, standing right at the intersection of skyrocketing AI compute demand and exponential rack-level heat dissipation, understanding and mastering AI has become a challenge I must tackle directly.
My initial idea was straightforward: could I collaborate with an AI agent to track AIDC industry intelligence more effectively? That is how this project started, originally under the working name dc-intelligence.
Soon, as agent capabilities improved, I realized we could do much more than aggregate industry news. The agent could pair with me to build interactive prototypes—and eventually, a full 3D thermal system simulation. As our collaboration expanded, I asked AI to register a domain and built this site to host the work.
How It's Built
Project A: Monthly Data Center Intelligence Report
- A wider radar: The agent continuously monitors public web sources—regulatory policies, capital investments, academic research, vendor equipment launches, and technical discussions across the industry. It gives me a broader, faster perspective on market dynamics than I could manage manually.
- Domain judgment over raw data: Pure information aggregation has little value on its own. I filter the noise through 15 years of engineering experience, cutting through marketing claims, highlighting physical constraints and contradictions, and distilling the data into a focused monthly assessment.
- From simple scripts to agents: The collection pipeline evolved from basic scrapers into agentic bots (such as Grok-based workflows) capable of managing context, maintaining memory, and cross-checking claims against primary sources. Down the road, as post-training costs decline and personal agent harness frameworks mature, I may develop a dedicated agent tailored specifically for AIDC.
Project B: 3D AIDC Solution Studio
Modern AIDC cooling as a whole is remarkably complex. Chip-level cold plates, secondary CDUs, primary chilled water loops, outdoor dry coolers or chillers, and hall-level airflow containment must all balance across tight temperature deltas, flow rates, and pressure drops.
When models began demonstrating strong 3D generation and spatial programming capabilities, I had an idea: could we build an interactive 3D tool for AIDC thermal design? It could serve as a hands-on learning tool for peers, as well as a visual sandbox for testing different cooling architectures.
Here is how the human-AI partnership actually works:
- I define the architecture: I design the system logic from scratch, sketch layouts, and establish the physical boundaries and fluid loops.
- AI assists with execution: It writes the Three.js and interface code, and searches public vendor documents for reference specifications.
- I guard the physical reality: This is the most crucial step. I review every capacity rating, fluid type, and temperature delta against real-world engineering standards, ensuring the model never conflates incompatible operating conditions or hallucinates physical impossibilities.
What's Next
This site remains an evolving project. As AI agents continue to advance and cooling requirements in AIDC grow ever more demanding, I will keep sharing technical insights from the field.
In my next post, I will share the unvarnished story of my "vibe coding" experience—how we divided the work, the engineering traps that only a domain specialist could catch, and the new prototypes currently taking shape.
Let's keep building.
从 ChatGPT-3 的横空出世，到 GPT-4、Gemini、Claude 接连迭代，大模型的能力一路狂飙。作为一个每天和实体工程打交道的工程师，我非常好奇：这个工具——或者说工作伙伴——的能力边界到底在哪里？
我做了 15 年暖通空调（HVAC）相关的工作，为各种不同的楼宇建筑，从工业、商业、酒店、办公到住宅。数据中心建筑的HVAC，也曾略有涉及，从电信机房的精密空调，互联网IDC机房，到二氧化碳两相高热密度液冷（10年前12kW/rack已经算当时的高热密度）。如今恰好站在“AI算力极速增长”与“rack物理散热指数增长”的交汇点上，深入了解并用好 AI，成了我必须面对的课题。
起初，我的想法很简单：能不能让 AI Agent 帮我更及时、更全面地捕获 AIDC 行业的动向？这就是这个项目最初叫 dc-intelligence 的原因。
但很快我发现，随着 Agent 工具的演进，它能帮我做的事情远不止抓取信息。它开始能和我搭档，从零写出数据中心热管理系统的交互原型，甚至做出一套完整的 3D 系统演示。既然能一起做的事情越来越多，我干脆让AI帮我注册了域名，把这个网站搭了起来。
它是怎么做出来的
项目一：数据中心行业情报月报（Monthly Intelligence Report）
- 更宽广的信息雷达：AI 负责在全网持续监测公开信息，覆盖监管政策、算力基建投资、前沿学术研究、厂商新发布的制冷设备，以及行业内部的技术讨论。它帮我建立起一个更新更广的市场视野。
- 用行业经验沉淀结论：单纯聚合公开信息意义有限。我把 15 年的暖通经验和物理常识放进去，剥离厂商营销话术，指出数据背后的约束条件和矛盾点，提炼成每月一期的深度研判。
- 工具本身的进化：底层抓取也从最初简单的脚本爬虫，逐步演进为具备上下文记忆、自主检索和交叉事实核查能力的 Agent（例如引入 Grok 等 Bot），信噪比得到了显著提升。等后面post training成本降下来，personal agent的harness 结构更成熟了，我可能也会开发一个AIDC 的专用Agent。
项目二：AIDC 3D 系统设计工作室（3D AIDC Studio）
现代 AIDC 的整个冷却系统非常复杂——芯片级冷板、二次侧 CDU、一次侧冷冻水、室外干冷器或冷机，再加上机房内的风液协同，涉及极其精细的温差、流量与压降平衡。
当看到大模型在 3D 渲染和空间代码生成上展现出惊人能力时，我萌生了一个想法：能不能做一款直观展示 AIDC 热管理运行机制的 3D 交互工具？它既能作为同行自学和培训的教具，也能帮工程师直观验证不同的系统架构。
这个模块的人机协作非常真实：
- 我出构想与架构：从零梳理系统回路、手绘草图、定义设备间的流体关系与物理规则。
- AI 负责落地与检索：协助编写 Three.js 和交互代码，并在公开资料中检索设备参考参数。
- 我做物理逻辑的守门人：这是最关键的一环。我用真实的工程经验严格审查数据真实性，绝不把不同测试工况下的冷量混为一谈，确保系统严格遵守热力学常识与工程规范。
写在后面
这是一个处于持续迭代中的项目。随着 AI Agent 的快速演进和 AIDC 市场的澎湃需求，我会继续在这个垂直领域输出一线洞察。
在下一篇文章中，我会专门聊聊我和 AI “vibe coding” 的真实经历——我们具体怎样分工、踩过哪些只有工程老兵能识破的坑，以及正在孵化的新点子。
人生没有终点，继续动手做点东西。