Our summary
Channel3 is building one catalogue of physical products gathered from many shops. Every product must land in the right category out of about 6,000, such as Bicycles or Coats and Jackets. Founding engineer Evan Fenster tested whether Jev, a TypeSafe AI that picks answers instead of writing text, could do this cheaply for tens of millions of products.
First the team walked down the category tree, keeping only the likeliest branches at each level, and a youth bicycle landed under motor vehicles. Asking Jev about all 6,000 categories reached 98.0 percent but cost $1.08 per product. The fix: a cheaper search shortlists 30 categories, then Jev picks one.
That setup reached 97.2 percent, up from 81.9 percent with the small model it replaced, at about $0.0005 per product. Jev did best on short lists, without training examples. It struggled to narrow huge lists step by step, and against cheap open models run in bulk it saved about half, not many times.
Key takeaways
- Shortlist first, then let Jev choose: Channel3's accuracy rose from 81.9% to 97.2% at roughly the same cost per product.
- Many cheap calls still add up: scoring all 6,000 categories cost $1.08 per product, about 2,000 times Channel3's production setup.
- Compare against your real alternative: next to cheap open models run in large batches, such as GLM-5.3-Flash, Channel3 found no meaningful saving.