Home / Case studies / Industry compliance document generation engine
Companies compile a full set of policies to national standards, moving from “hire a consultancy” to “generate it yourself in 15 minutes”.
To pass work safety standardisation review, a company must compile a full set of policies, emergency plans and registers.
The traditional route is to hire a consultancy: around CNY 50,000–100,000, 15–20 days. That is a heavy burden for smaller companies, and what you get is a generic template detached from your actual organisational structure and equipment list.
The essence of this step is not 'writing articles' but 'translating national standard clauses into documents a company can actually execute'.
In-house vertical LLM application: national standard clauses are structured into a knowledge base, document sets are broken into generatable templates, and once the company fills in its basic information the whole document set is generated in one pass.
The knowledge base covers two national standard systems: a full system for large and medium enterprises and a streamlined one for small businesses.
Results can be edited online and exported to Word, so companies can revise them to their own situation and use them straight away.
Beyond standard-compliance documents, we also provide tools for hazard identification, risk notification cards, emergency plans, safety signage and training plans.
The table below lists this project's technical elements so clients can verify them independently.
Project stories without constraints are fiction. This section sets out four constraints: time, site, staffing and compliance.
These constraints directly shape the solution; every problem under 'Key challenges' below is forced out by one of them.
It is not “we worked hard”; it is “what the problem looked like, how we solved it, and what it cost”.
This section is for peers. If you are doing something similar, these points will save you a few detours.
Any scenario that “issues documents to a standard” must first break the standard into structured fields; otherwise references will go wrong.
When the format is strict, do not let the model decide the structure.
It removes compliance concerns but raises the hardware bar. Choose according to each client's actual situation, not with a blanket rule.
Talking only about successes is unconvincing. This section covers what we got wrong or did too late.
Doing it again, we would build the industry template library earlier. As it is, every industry needs templates configured from scratch; a generic template engine abstracted at the start would make adding industries much faster.
We started evaluating generation quality too late. Early on we judged 'does it look right' by eye with no quantitative metrics, so tuning was guided by gut feel. It improved only after we built an evaluation set.
Mobile was not built at all. We assumed such tools were only used on PC, but many small business owners want to check results on their phones and could only be given a read-only page.