Who was stuck
A builder in NSW checking one wall, one stair or one balustrade against the National Construction Code. The answer exists, but it is buried somewhere in hundreds of pages of NCC 2025 and the ABCB Housing Provisions, spread across cross-references that point at other cross-references.
Why it happens
The NCC is huge and hard to search. One question can depend on the building class, the climate zone and the state, and each of those changes which clause applies. So builders end up paying a consultant or guessing.
What we built
Vecton is a compliance assistant for the NCC. You ask a question in plain English and it answers with the clause attached, so you can check the source instead of trusting the summary.
- Retrieval over NCC 2025 Volume Two and the ABCB Housing Provisions, mixing meaning-based and keyword search.
- Every answer cites the clause it came from, filtered by building class, climate zone and jurisdiction.
- My role: customer liaison and programmer in an eight-person capstone team, working directly with the client on what "traceable" had to mean.
Try Vecton live and ask it about a clause. It's hosted on a free server that sleeps when no one's using it, so the first load can take up to a minute.
The call we made
When Vecton isn't sure, it says so and doesn't answer. We could have let it have a go at every question, and it would have looked smarter in a demo. But a builder who trusts a wrong answer can fail an inspection, or worse. The client told us early that being right mattered more than answering everything, so we made declining the default whenever it couldn't find the clause.
Did it work
The client's top requirement was correctness over coverage: a wrong answer is worse than no answer. So we tested it on 15 fixed questions, run on the live system, with a script that checks every quote and number against the clause it cites.
- 11 of 11 in-scope questions cited the expected clause, and every answer was backed by its source.
- 5 of 5 out-of-scope questions were declined, with 0 invented clauses or facts.
- 270 NCC figure descriptions checked by hand against the drawings.
- 570 automated tests, with 94% coverage on both the backend and the frontend, running on every pull request.