31 batches and 1,642 creatives went into one catalogued library. Then the live library quietly showed 7 of them, and the thumbs-up learning loop starved after 212 ratings. What a creative registry needs beyond “put everything in one place”.
Read More
A product launch for us touches about twenty surfaces. We turned it into a durable publication run with preconditions and a post-write read-back. 237 runs later, a quarter failed, and almost none of the failures were in our logic. Here is what failed, the six launch bugs that each became a hard rule
Read More
A Sunday job crawls our marketplace reviews, emails a digest and turns new five-star reviews into social posts. For a month it reported “0 new reviews” while customers wrote 46. Here is the data model, the three bugs, the gate that rejected seven five-star reviews, and the alert I should have built
Read More
Nine weeks of a weekly SEO job that recommends exactly three fixes: the real numbers, the title fix it called a win while clicks fell, the three weeks it repeated itself, and the verdict rule I would build first next time.
Read More
In one September week our daily ad spend went from a quarter of the budget’s daily rate to 4.6 times it, with nothing broken. Here is the guard I wrote afterwards, the ranking code it runs every three hours, the cents-versus-dollars bug that briefly multiplied our account cap by 100, and the weekly
Read More
We push one price to seven sales channels from a single table. An audit found zero drift on our own store and 130 problems on the marketplaces, most of them leftover deals. Then a price edit quietly reverted itself and every channel reported success on the wrong number. The model, the log, and the t
Read More
Outscraper as the Google truth layer for a national shop directory: places snapshots, one-time review texts, Directory Insights screens, first-run vs weekly cost, disclosed referral.
Read More
Three months of running a retail company with an AI team: 12,132 logged agent tasks, 158 scheduled jobs, and a human task queue where 611 of 758 items are still open. The bottleneck was never the agents. Here is the data, the design mistakes, and what I would build first.
Read More
Concrete screenshots from running an AI ops team in Cursor for ANDAFACIL & MandaFacil — weekly review digests, creatives Approve & post, SEO→3 actions, SofIA blog drafts — plus Agentic Company (MIT).
Read More
AI-powered tools like Cursor make market research & data analysis effortless. Learn how to extract insights, estimate revenue, and analyze competitors—no coding required!
Read More