A growth system for NewWay, a womenswear manufacturer with $100M in yearly exports. An autonomous agent researches fashion brands, designs each one a custom lookbook, and writes email real buyers answer. I also rebuilt the brand's English website.
What I did
Agent architecture
Prompt and judge design
Pipeline code
Daily ops
Client website rebuild
2,343
Brands researched into the pool
389
Custom lookbook PDFs generated
51
Consecutive daily runs
32
Replies · Eileen Fisher, Vince, Revolve
The problem
Generic manufacturer email gets deleted on sight. Real BD works because a human studies the brand first. That hour per brand is why nobody does it at volume.
Can an agent do the hour of homework for 22 brands a day, to a standard a skeptical fashion buyer can't tell from human work?
What I built
- The full agent system, mapped in the chart above: signal scanners, a tiered allocator, brand curators, three AI judges, a sender, and a reply classifier.
- The lookbook generator, a pipeline that turns catalog picks into a branded PDF per prospect. 389 produced.
- The quality bar, nine iron rules and judge prompts, each written from a documented failure.
- The client website, newway-apparel.com, rebuilt with every old Google indexed URL preserved and its AI search readiness lifted from 63 to about 71.
Decisions that mattered
- Nine iron rules, each tied to a documented failure. v3 never logged its failures and output collapsed. v4 records everything.
- A kill switch in one flag. The whole system reverts to drafts with zero code change.
- The pitch inverts by brand size. "MOQ from 100" wins small brands and loses big ones, so large prospects hear the Zara and H&M production story instead.
- A lessons file 5,000 lines long. Every agent reads it before starting.
Real brands answer email a machine wrote: 32 replies across some 440 tracked sends, Eileen Fisher and Vince among them. One is in active negotiation.