结束时，有东西真的跑起来了。这不是课堂。
一个团队、一个真实应用场景、一种强化形式，由每天构建这类系统的工程师带教。结束时会有一个基于贵司数据的可运行原型——以及一个知道它为什么能跑起来的团队。
- 围绕贵司真实案例
- 由工程师带教
- 结束时交付原型
The gap between understanding and being able to build
You can sit through ten hours of training and still be unable to start a project. The format has a lot to do with it.
Understanding is not building
The concepts are clear in the room. In front of a real corpus, with real access rights and real constraints, nothing looks like the example any more.
Momentum fades
Months pass between the training and the first project. Decisions dilute, people change priorities, and nothing starts.
Nobody dares fail
With no space to fail without consequence, teams only try what they already master. So they learn nothing new.
How it runs
The case is chosen before the start, on real data. The format does not work on a fictional subject.
Choosing the case
Before the start we pick a real, useful use case that is achievable within the format — narrow rather than ambitious. Too broad a case produces an unfinished demo, which is demoralising and useless.
Preparing data and access
Data is prepared and access is opened before day one. Spending the first two days waiting for a permission is the most common way to waste this format.
Foundations
The concepts needed for the chosen case, and only those: what a model does well, what it gets wrong, how to wire a source, how to verify an answer. A minority of the time.
Coached building
The team builds, with an engineer available throughout. Dead ends are embraced: understanding why an approach fails on your corpus is worth more than a successful demo on a hand-picked one.
Presentation and next steps
The prototype is presented, its limits documented, and the question asked frankly: industrialise it, rebuild it differently, or drop it. All three answers are acceptable.
Cases that work well in this format
The common criterion: a narrow perimeter and data available on day one.
Assistant on a bounded corpus
A specific document base — one department's procedures, a product catalogue, a regulatory set — made queryable.
Document extraction
Pulling structured fields out of a recurring document type and checking them.
Dataset analysis
Querying a file or database in plain language and verifying the results are correct.
Task automation
A simple end-to-end sequence, with a human sign-off point.
Document comparison
Detecting the real differences between two versions or two contracts.
Evaluating an approach
Deciding, on your data, between two ways of solving the same problem.
What the team takes away
- A working prototype on your case and your real data
- The code and configuration, documented and reusable
- The prototype's measured limits, quantified rather than felt
- A frank recommendation: industrialise, rebuild differently, or stop
- A team able to start again on their own on a neighbouring case
常见问题
A small team mixing business and technical profiles. The format loses all value when nobody knows the domain, or nobody can build.
Not everyone. You need at least one technically comfortable person on the team, and people who genuinely know the business.
Yes, that is the condition. On fictional data the format proves nothing and teaches nothing useful.
As a prototype, yes. In production, no: rights management, traceability and robustness are missing. Industrialisation is what brings them.
That is a result, and often the most profitable one: knowing within days that a path does not hold avoids a six-month project.
Both work. On site the team dynamic is better; online the logistics are simpler for distributed teams.
本方案不做什么
A bootcamp prototype is not a production system. It has neither the rights management, nor the traceability, nor the robustness of a deployed system — and confusing it with a finished product is the best way to disappoint everyone three months later. The format exists to decide quickly and to build a team's capability, not to save the cost of industrialisation.
Which question deserves to be settled quickly?
If something has been going round in circles in your meetings for six months, this format answers it in a few days, on your real data.