The application we kept rebuilding until rebuilding was the method.
Product Campaign Manager was our proving ground: an autonomous marketing-orchestration platform we built, tested, tore down, and built again — generation after generation — while the delivery discipline underneath it hardened into the lifecycle we run everything on today.
Every figure on this page is drawn from the project’s own records — its estimates, its commit history, its test logs. The repository is private; the numbers are not.
The first build — the numbers that started it
The original PCM build is the case study we’ve published since: requirements and discovery in 6 hours through Alchemist; a functional, cloud-deployed system from the crew in 9 hours of build execution.
1,086 test scripts tied to acceptance criteria
4 weeks of calendar against a 50-week plan
$250K actual cost against a $2.27M plan — roughly $2M unspent
a team of 3–4 against a planned 9
Those numbers were the beginning, not the point. What we did next is the part this page exists for.
Then we kept rebuilding it — on purpose
Most teams treat a working system as something to protect. We treated PCM as something to regenerate — the record in front of us is generation six of the same application, rebuilt from evolving requirements each time the platform’s capability grew. A rebuild wasn’t a failure to preserve the old version. It was the cheapest way to give the application everything the discipline had learned since.
The final generation’s scope, estimated the traditional way at kickoff — the estimate is written into the project record, story points times hours:
2,205 story points × 4 hours = 8,820 estimated development hours
≈ 55 team-weeks. More than a year of traditional delivery — for one generation.
The actual record: 1,068 commits between April 10 and May 14 — about five weeks of calendar, covering the crew build (two crew runs, A and B) and the entire human final mile.
The final mile, run like a lab notebook
The Away Team work — taking crew output to production-grade — ran as nineteen documented episodes, and the discipline shows in the log itself. Each episode closed the same way: a layered test pass with its count printed, and a numbered lesson added to the project’s lessons log. Entries from the record, verbatim:
Notice what the fixes are: surgical. One URL. One scaffold. A test assertion. When a build executes against hyper-defined requirements, the final mile isn’t archaeology — it’s alignment, and every episode proves it with a test count instead of a feeling.
By the log’s close the lessons numbered sixty-one. They didn’t retire with the product — they ride in every build that came after. That transfer is the whole reason the log exists.
Where PCM is now — because a receipts page doesn’t get to skip this part
PCM is retired. Not because it failed — because the capabilities it orchestrated moved into the platform itself, and keeping a separate application alive for sentiment is exactly the kind of cost the discipline exists to catch. We stopped at episode nineteen with the continuation report already written for a twentieth that strategy made unnecessary.
What survived is everything this page is made of: the delivery figures, the sixty-one lessons, and the method — proven here first — that now carries our CRM, a commercial client’s public site, and a video game a kid designed.
Software you can afford to retire is the same thing as software you could afford to build. That’s the receipt most delivery organizations can never print.
The discipline PCM proved is the one we’ll run on your scope. Bring us the estimate you don’t believe.
Alchemist AI Pro™ is free to use — alchemistaipro.com
Figures verified against the project’s own repository records: the deterministic estimate written at kickoff, the commit history (1,068 commits, April 10 – May 14), the per-episode test logs, and the numbered lessons ledger. The repository is private; the original build’s case-study figures have been published since 2026.