Send the quote.
Before you've left
the driveway.

A UK sole trader quotes from a photo of a wall, a voice note in the van, a half-remembered chat at the door. Then the evening goes on writing it up. And the job goes to whoever replies first, not whoever is best.

The failure moment is specific: turning mess into a priced, customer-ready quote while still on site. Nothing cheap rehearses that moment. That is the whole product.

The tools are split into two halves: prepare and apply. The model can only prepare a change, which puts it on hold and asks the tradesperson to confirm it in their own words. The change is only actually applied once the tradesperson's reply has been checked and accepted as a genuine confirmation, and the model has no way to make that final step happen itself. Nothing is ever overwritten: every change is saved as a new entry, with the previous one simply marked as no longer current, and the event log is locked down so that any attempt to edit or delete a past entry is automatically blocked.

Gate two is my favourite thing in the codebase, because it exists as the result of being caught. An audit found that any part of the system could simply mark a piece of information as "confirmed" without ever checking it against where that information actually came from, which meant the gate did nothing to catch data the AI had merely guessed rather than verified. The fix closed that backdoor: now something can only be marked confirmed by tracing it back to its real source, never by just declaring it so. The comment recording that fix is still in the file, with the date.

The near miss that justifies keeping the arithmetic in code at all, never the model: an early version silently dropped two labourers from a quote entirely. My own log at the time: "it's missing my two mates, labouring for £180 a day for 5 days." £1,800 of real labour, gone, caught before it reached a customer.

Anyone can call the same model I do. A sole trader's voice note says "chuck in some T&E and a couple of double sockets." A generic assistant hears a sentence and writes something plausible and useless; this one already knows the trade.

What a fresh competitor cannot buy on day one is the layer above it. Every approval and every edit is a signal, weighted into that user's profile, so the baseline improves and the edits shrink. The design bet was that a competitor starting from scratch could not shortcut six months of accumulated signal, a bet the project never ran long enough to test. The ladder below is how one correction was meant to become product-wide confidence without ever skipping the human at the top.

Signal comes from the highest-quality source available: what the tradesperson changed in a draft before sending it, plus their reactions, voice and replies. Each time the same pattern shows up again, the system's confidence in it grows by a fixed tenth of a point, though it levels off at a ceiling rather than climbing forever. Repeats are recognised automatically within a seven-day window, so the same pattern can never be raised twice.

The rung that matters is the third one: the system asks the tradesperson to confirm or dismiss what it thinks it has learned, as a first-class tool call. And the two brakes at the bottom are what I would want a reviewer to notice: regulatory patterns are skipped outright, and nothing anywhere in this ladder writes product configuration without a person.

The standing rule for cross-model review was "Agreeing is a failed review." The rule that makes it usable rather than noisy is a second one: every finding must cite a verbatim file and line before it counts. A finding without a code anchor is not a finding. Every finding is triaged as confirmed, partly debunked or needs-verify before it counts, which routinely reclassifies some as false positives, and occasionally shows the bug is worse than the discoverer claimed.

I am showing this deliberately, and not as a roadmap. It is the shape of what the grind actually bought. The real return on Van Brain was never going to be Van Brain. It was learning the stack properly, learning how to build with AI agents rather than read about it, learning to test an idea and kill it, and learning to hold a long-term strategy loosely enough to abandon it. That meta-learning is the thing I took out of it, and it is why the graveyard is worth looking at rather than hiding.

Two tradespeople entered the funnel; only one, Harvey, ever sent it to a real customer. He paid on 11 June, on a discount steep enough that no money ever actually changed hands, and was never activated: a mismatched security credential between two systems left the activation silently dead for twelve days, then an old, leftover part of the sign-up process corrupted his profile and crashed the system. The recovery took a full session with multiple AI agents working together, and thirteen rounds of code fixes, numbered 689 to 703. Then came the pilot's actual killer: Harvey forwarded a client's itemised email brief for a garden office, and the system smoothed the detailed brief into a generic summary instead of keeping the itemised detail intact, the customer's own detail lost.

Charlie, the second tradesman, never completed a real pilot session; his sharpest use was as a founder-driven dogfood account, and it caught the worst bug in the project on its own: ten explicitly priced materials silently collapsing into six generic categories at quantity zero, a £135.60 wrong quote on an £800 job. Caught before it ever reached a customer, which is the whole point of building the scorecards in the first place.

The deeper problem was the order I did things in. My own audit flagged demand as the number one risk on the fourteenth of May, and I kept building for three more weeks. The single largest category of merged work was not any code surface: it was documents and strategy, at just under a third of everything. The one cheap way to test whether anyone wanted this, a waitlist and a few pounds of ads, sat at the top of my own morning briefing as the highest-leverage unclaimed task for twenty consecutive days and I never ran it.

The honest verdict on the product idea came out of the testing. My builders did not want a better chat interface. They wanted a magic machine, materially smarter than anything generally available, and the only route to that is a team of agents sitting on proprietary pricing data. Which means merchant partnerships, a large customer base to generate the data in the first place, and a great deal more tooling in a domain that is not my profession. Plenty of people would have liked the invoicing app. Liking an invoicing app is not a moat.

The commercial figures stay off this public page by choice. Every visual on it is public-safe demo data or the project's own test fixtures, never a live customer export.