Event
Watch again: building a close application in forty minutes
Forty minutes, one paragraph describing a month-end close, and the application that came back — corrected twice on the call, and rebuilt from the corrected…
By Sofia Marchetti · Field engineering · 17 June 2026 · 4 min read
Forty minutes, one paragraph, and a working application by the end of it — with the two corrections it needed made on the call rather than edited out of the recording.
The paragraph was a real one, contributed by a group controller who has run a four-entity close for several years. It described nine ledgers, the order they have to be reconciled in, what counts as matched, what has to be explained instead, and what the audit committee expects to read. It named no technology, which is the point.
The running order
The session ran in five parts, and the proportions are the interesting bit — more than a third of it was correcting.
- Describe — reading the paragraph out loud and marking the places it was ambiguous. Two were found before anything was built.
- Build — the four documents derived in order, and the application generated from them.
- Correct — the part that took longest, and the part worth watching.
- Rebuild — regenerating from the corrected description rather than patching the code.
- Call it — submitting an operation, polling the job, reading the pack out of the store.
The two corrections
Both are the ordinary kind, which is why they were worth doing live.
The first was an intercompany rule that was too generous. The description said balances between entities should net; it did not say that a difference under a threshold could be written off and a difference over it had to be explained. The application had reasonably assumed the first behaviour for everything, and the controller spotted it in the first result.
The second was a cut-off that was a day out. The paragraph said "at month end", and the group's actual practice takes rates at the close of the last business day, which in a month ending on a Sunday is a Friday. This is exactly the kind of thing that is obvious to the person who does the work and invisible to everybody else.
Neither correction was made by editing code. Both were made by editing the description and regenerating — which took under two minutes and is the behaviour the session existed to show.
What was deliberately not shown
Three things, and they were named on the call rather than skipped quietly.
Nothing scheduled the run. The operation was submitted by hand, and in production it would be submitted by the customer's own scheduler when their source systems land. There is no scheduling page in the product and none was implied.
The figures on screen were invented. The ledgers, the entity names and every number in the result were made up for the session. What was real was the shape: the operations, the job id, the poll, the output written under the organisation's own path.
No compliance artefact was produced. Somebody asked whether the pack came with a record of who ran what, against which data. The answer given was the honest one: not from the platform, today. Where a process needs that, it is built as part of the application's own outputs.
Watching it
The recording is unlisted rather than gated. There is no form in front of it, no email required, and nothing that follows you around afterwards — which is a decision about how we would like to be marketed to, applied to ourselves.
If you want the paragraph that started it, it is in the description under the video, unedited. It is shorter than you would expect, and that is most of the lesson.
The figures in this piece — including the ones in the panels — are illustrative. They show what a result of this shape looks like, not a measurement of anything that happened.
Read next
Save your place
What happens next
Thank you. We will write once, when it's ready.