Ground Truth · case notes from AI in real businesses真实企业的 AI 落地案例

AI that survives contact with a real business.让 AI 在真实的生意里活下来。

I’m James Wang. I put AI into law offices, restaurants, schools and job sites, then write down what worked, what broke, and why.我是 James Wang。我把 AI 做进律所、餐厅、学校和施工现场，再把哪些有效、哪些出了问题、为什么，一条条写下来。

Forward-deployed engineer. I work inside your operation, ship the system with your team, and stay until your people run it without me.FDE，前线部署工程师。我进到你的业务现场，和你的团队一起把系统做出来、跑起来，直到不需要我也能运转。

I run a business我是企业主 → readiness check先做自测 I want to learn我想学 AI → case notes看案例

trace · lawyerparaillustrative示意

09:47:12intakeclient message · matter NV-218
09:47:12redact2 names → "the client" · 1 file withheld
09:47:13retrievefirm policy · fee schedule v3
09:47:14modeldraft status update · provider: swappable
09:47:14tierClient Comms = T2 · send blocked
09:47:15queuewaiting for partner approval
09:52:30approvepartner edited 2 lines · approved
09:52:30restoreclient name re-inserted
09:52:31senddelivered · logged with approver

Gold lines: a person decides. The model never sees the client’s name.金色行：由人来决定。模型从头到尾看不到客户的名字。

01 · Case library案例库

Every case starts with something an owner actually said.每个案例，都从老板说的一句真话开始。

For each one: the problem in plain words, what shipped, the guardrail that made people trust it, and the stack underneath. Filter by your industry.每个案例都写清楚四件事：问题本身、最后交付了什么、让人敢用的那道防线、底层用了什么技术。可以按行业筛选。

02 · How an engagement runs合作流程

One workflow, shipped to real users, then handed over.先做一个流程，上线给真实用户，再交给你的团队。

Most AI projects stall because they start too wide. I start with the one task that costs you the most time, and I don’t leave until someone on your team owns it.很多 AI 项目卡住，是因为一开始铺得太大。我从最耗你时间的那一件事入手，直到你团队里有人能接手才离开。

Step 1

Sit in进现场

1 to 2 days1 到 2 天

I watch the work, collect real examples, and find where hours and customers are lost.看你们实际怎么干活，收集真实样本，找出时间和客户在哪里流失。

You get:交付： a one-page problem brief一页问题说明

Step 2

Map the risk画出风险

about 1 week约 1 周

Pick one workflow. Decide what the AI may read, write and send, and which number will prove it worked.选定一个流程。定清楚 AI 能读什么、写什么、发什么，以及用哪个数字证明有效。

You get:交付： permissions and a success metric权限表和成功指标

Step 3

Pilot小范围试点

2 to 4 weeks2 到 4 周

Ship a narrow version to real users with human approval switched on. Fix what they hit.把收窄后的版本交给真实用户，保留人工审批，用到哪里不顺就修哪里。

You get:交付： a working system and its logs能用的系统和完整日志

Step 4

Harden and hand over加固并交接

1 to 2 weeks1 到 2 周

Tests, audit logs, fallbacks for when the model is down, and training for the person who will own it.补齐测试、审计日志、模型宕机时的兜底方案，并培训以后负责它的人。

You get:交付： docs, runbook and an owner文档、运维手册和负责人

The AI drafts. A person approves.AI 起草，人来拍板。Anything that reaches a customer goes past someone on your team first.凡是发给客户的内容，都先经过你团队里的人。

Your data stays yours.数据始终是你的。Sensitive fields stay out of the model, and everything is exportable.敏感信息不进模型，所有数据随时可导出。

No lock-in.不绑定。Mainstream stack, swappable model providers, and code your next engineer can read.主流技术栈，模型供应商可替换，下一个工程师也看得懂代码。

03 · For business owners给企业主

Is your business ready for an AI pilot?你的生意，适合做 AI 试点吗？

Six questions, about a minute. Your answers stay in this browser. The result tells you where I would start.六个问题，大约一分钟。答案只保存在你自己的浏览器里，结果会告诉你我会从哪里入手。

04 · For learners and students给想学 AI 的人

Learn the part tutorials skip.学教程里不讲的那部分。

Prompting is a small part of the job. The rest is understanding a business, drawing the boundaries, and keeping the system alive after launch. That is what I write about.写提示词只是工作里很小的一部分。剩下的，是读懂一门生意、划清 AI 的边界、上线以后让系统活下去。我写的就是这些。

Case notes案例拆解

One real deployment per episode

每集拆一个真实项目

The problem, the architecture, the guardrail, and the trade-off I made.

问题、架构、防线，以及我做了什么取舍。

Diagrams架构图

How the pieces connect

各个部件怎么连起来

RAG, permission tiers, voice hand-offs and model fallbacks, drawn plainly.

RAG、权限分级、语音转人工、模型兜底，用最直白的图讲清楚。

Mistakes踩过的坑

What broke and what I changed

哪里出了问题，我改了什么

The lessons that only show up after real users arrive.

只有真实用户进来之后，才会暴露的问题。

Path成长路径

From developer to forward-deployed engineer

从开发者到 FDE

The skills I’d learn again, in the order I’d learn them.

如果重来一次，我会按什么顺序学哪些本事。

Season 1: Built in Production第一季：Built in Production

05 · About关于我

I started out teaching web design in Harbin and building websites for American companies from the other side of the world. Twenty-three years later, I build AI for the law office and the restaurant down the road.我从哈尔滨的一间教室起步，教网页设计，也隔着半个地球给美国公司做网站。二十三年后，我在给街角的律所和餐厅做 AI。

Along the way I led a 15-person engineering team, ran booking systems behind a load balancer rated at 10K requests per second, and built manufacturing software under FDA rules. Today I’m CTO of AAWEA, a nonprofit for entrepreneurs, and a senior engineer at QIAGEN.这一路上，我带过 15 人的研发团队，做过承载每秒上万请求的酒店预订系统，也做过符合 FDA 规范的生产管理软件。现在我是创业者公益组织 AAWEA 的 CTO，同时在 QIAGEN 做高级工程师。

1999Instructor, Harbin University of Science & Technology哈尔滨理工大学，网页设计讲师 2003Web developer to tech lead, Harbin哈尔滨，从开发做到技术负责人 2013Moved to the U.S.; led a team from zero to 15来到美国，团队从 0 带到 15 人 2021Senior Software Engineer, QIAGENQIAGEN 高级软件工程师 2025CTO, AAWEA; 11 AI platforms sinceAAWEA CTO，至今上线 11 个 AI 平台

Work with me找我合作

Bring me one problem, not a strategy.带一个具体问题来，不用带战略。

Email me three things and I’ll tell you honestly whether AI is the right tool:发邮件告诉我这三件事，我会如实告诉你 AI 是不是合适的工具：

What your business does, and how many people work there你做什么生意，团队有多少人
The task that eats the most hours every week每周最耗时间的那件事
What a good result would look like in numbers做成了，用什么数字来衡量

[email protected]

LinkedIn · jameswangusa ↗ · Resume & casebook ↗