At Viya Constructions, three to five people who know construction build the company’s software with AI coding tools. What they couldn’t tell was whether it was safe to run a company on. This is how we closed that gap.
[01]Where it standsSelf-Build for Viya
Every result below was confirmed by checking the actual code, not by taking a note’s word for it.
Findings are still being worked through, including the feature problems from the late check. The first phase was a floor, not a finish line.
The know-how stayed inside the company, the team kept shipping at its own pace, and now someone can say, with evidence, whether it’s safe to run the business on.
[02]Where it startedSelf-Build for Viya
Viya’s own people build the software the business runs on. Someone who knows exactly how a rebar schedule is calculated, or how a punch list moves between an architect and a contractor, can build the software for it directly.
From the outside, the software looked finished. People logged in and the screens loaded. What you couldn’t see was whether one customer’s data could reach another’s, whether a former employee’s access really ended, or whether something the screen said was saved had actually been saved.
[03]How it runsSelf-Build for Viya
A six-step loop, every week. The step that matters most is the last one.
The team builds on its own priorities. For payments, logins or sensitive data, they check the approach with us first. Who: Viya, with us on call.
Everything new is checked for security, customer separation, correctness, tests and release setup. Who: us, weekly.
Same day, same format. Each finding with how serious it is, why it happened, what it risks and how to fix it. Who: us.
Questions on the report get answered, so nobody starts fixing from a misunderstanding. Who: both.
The team fixes things in the order the report sets. The next review checks the actual fix, not a note saying it’s done. Who: Viya.
Anything that showed up twice gets an automatic check. This is the step that makes each week cheaper. Who: us.
We’re on call for the decisions that are expensive to get wrong. A review can’t help with these, because it comes afterward. The AI tool offers several designs with the same confidence, and picking the safe one takes experience.
For a builder without an engineering background, this is where they’re most on their own. It’s also the cheapest hour in the whole arrangement.
[04]The blind spotSelf-Build for Viya
The weekly review asks whether the code is safe, separated and tested. It doesn’t ask whether a feature does what its screen says.
No high-risk findings. The best result so far.
Including a drawing upload that said “success” and never sent the file anywhere.
Both were true because they measured different things. Finding that out was the most useful result of the work so far. It changed what we check, not just how often. We now check both.
[05]More detailSelf-Build for Viya
Check everything already built, rank whatever could stop it running as a real business system, and explain the cause and the fix for each. The first review is the starting point. After that, it’s a weekly check.
Payroll, and a rebar schedule tool with paid access per project, approval for offline payments, and export. Built to the standard from the first line, not fixed afterward.
Code and hosting moved off personal accounts and outside platforms onto accounts the business owns, with access that doesn’t depend on any one person.
Turn repeat problems into rules that reach the team while they build, and be on call for decisions that are expensive to get wrong.
“An app” turned out to be several: one platform serving staff, homeowners, contractors and architects; an admin app split out with its own logins; and the two new modules we built. Working out the real size of the job is always the first step.
One pattern showed up so often it’s closer to a law. A problem gets found and properly fixed. Then something new is built, like a new type of user or a second data importer, and the same problem is there, untouched. In the third review, eight of the twelve new findings were exactly this.
The fixes got much better. Carrying each fix to every place it belonged did not, because that depended on memory. This is how AI-assisted building fails: not randomly, but predictably, on whatever was built last.
It’s also why rules the computer enforces held every single time, and rules that relied on memory held only until the week got busy.
We list roles, not names, because the setup is what you can copy.
Findings arrive faster than one person can fix them. Knowing that early is better than finding out in month three.
If your people are already building with AI tools, the question isn’t whether it works. It’s whether anyone can tell you it’s safe.