I have one Amazon review and no client case studies. Rather than write vague quotes to fill the gap, here is what I can actually prove: what each system cost to build and run, logged per call at the time of the call, and what it produced. Including the parts that did not work, which is the half nobody publishes.
Costs were written per API call as the call happened, with real token counts. That makes these receipts rather than estimates, and it is the reason I am willing to print them next to my own failures.
Eleven cents per thousand words. Essays, short stories and columns, drafted, voice-scrubbed and machine-rated before a human ever saw them.
Source: per-run generation ledger, cost and word fields summed over the runs array0.073 cents each. Written per call at call time with real token counts, which makes it a receipt file rather than an estimate.
Source: per-call token log, 823 calls between 2026-05-28 and 2026-07-09445 scraper runs and 13.79 hours of compute. 35,564 raw records collected, 15,295 passed the quality bar. No language model anywhere in that pipeline, which is why it cost nothing.
Source: the 15 listings files, the scraper logs, and 102 run summariesPlain string interpolation, no API library anywhere in it. The honest caveat, printed here because it belongs next to the number: 600 of those pages come from 20 unique bodies. Real surface area, not 600 distinct pieces of writing.
Source: the two page-generation engines and their output directoriesEvery one of these is measured, sourced and unflattering. They are here because a cost receipt is only believable next to the record of what the cheap output was actually worth.
Not sampled. Counted. Read one at a time they were fine, which is exactly why I did not notice for a month.
Asked for a name with no further instruction, the model returned something plausible, and plausible has a shape. A death appeared in 69% of those stories, a mother in 75%, grief in 52%.
The best single post got two interactions. By then the writing had genuinely improved, which is the uncomfortable part: quality was never the bottleneck, distribution was.
Measured on my own output. It is why the current pipeline renders text in CSS and screenshots it instead of asking a model to draw letters.
No client testimonials. No case studies. No revenue screenshots. The AI Workspace Setup launched this month and the first three clients are discounted specifically to buy those, with permission agreed before the work starts rather than asked for afterwards. When they exist they will appear on this page with the before, the after, and the counted result. Until then this is what I have, and everything on the projects page is clickable.
Fifteen live directory sites, thirteen games, a CMS, an email engine and a checkout, all built solo with AI and all running right now. Nothing on this site is a mockup.
See everything live →