A phone scan of a room becomes a field plan, and the math stays in code

One project can mean three to twenty invoices across about five trades, and the reference job we worked from had four change orders on top of that. The person who described the invoicing situation to us called it a cluster, and the detail behind that word is what makes it real: multiple owners of the process over time, no clean handoff between them, and every job type carrying a different milestone structure.

The first ask was not automation. It was visibility. One place to see what has been sent, what is outstanding, what is overdue, and who has already been messaged. That is not a glamorous request and it is the correct one, because you cannot automate a process nobody can currently see.

The estimating side had its own version of the same waste. The measurements already existed. Somebody had scanned the rooms with a phone. Then a person spent one to three hours per job turning those measurements into line items by hand, room by room, trade by trade, and on a large project a single missed line costs thousands. The data was there. Nothing carried it forward.

And the collections story is the one worth sitting with. There was $151,000 past due on a single client, across more than ten invoices. The job management platform they run has automatic payment reminders built in, and those reminders were switched off. Not because nobody thought of it. Because they are context-blind: they will chase a client who paid yesterday, or one who is mid-dispute over a change order, and every one of those costs a relationship. So chasing went back to being manual, which means it did not happen.

We built the field-plan parser first. A phone LIDAR scan becomes a structured field plan: measurements, linear footage, cabinet runs, and the scope breakdown, drafted for review. It scores 100 out of 100 on its golden eval, which is a fixed set of real cases with known-correct answers that runs on every change, and it is live in production. An earlier pass scored 96.6, and the gap between those two numbers is about a week of finding out what the model got quietly wrong on real jobs.

The measurements that matter got included because we were told to include them, not because we designed them in. Cabinet and countertop measurements needed to appear on the field-plan summary even if somebody typed them in by hand at first. That is the kind of detail you only get by asking the person who does the work.

On the billing side, the structure is data rather than logic. 30 / 30 / 35 / 5 for a bathroom, percent complete for something else, and whatever the bank dictated on a financed job. A new job type is a configuration, not a rebuild. Every dollar in it is computed by code, never by a model, which is the same rule that governs the Pearl Street engine and for the same reason.

Collections is being built to be context-aware, which is the entire difference between it and the feature that already exists and got turned off. It knows what was sent, who replied, and what is contested. Nobody gets chased for an invoice they paid yesterday. The system's job is to tell a person who to call and why, and the call stays a person's.

The parser is live and scoring 100 out of 100 on its golden eval. The person reviewing a scope is checking a near-complete draft rather than assembling one from a blank page, which is the difference between a fifteen minute review and three hours of line-by-line work. The estimated time recovery across the operation is roughly 26 to 38 hours a month, and we are labeling that an estimate because that is what it is.

The $151,000 past due did not get collected by software. It got surfaced, ranked, and turned into a list somebody could actually work, which is the part that was missing. The invoicing and collections modules are in build.

The reference job behind all of this was a real kitchen and flooring remodel with a master estimate of $56,867.50 and fourteen real PDFs. We are not going to tell you whose house it was, and the customer names inside that job stay local and out of every repository we touch. That is not a policy paragraph. It is why the numbers on this page are the ones we are willing to show you.