Place products into a store category tree
Jev walks a product description down a category tree one typed choice at a time, replacing a slower agent loop.
Author's figure: Author reports mean latency per product fell from 9.62 s to 1.38 s and model calls from 7.22 to 3.18, with 36 of 50 paths matching the agent's.. Reported by the builder, not measured by us.
The use case
Online sellers and catalog teams must file each product under the right leaf of a taxonomy such as Shopify's or Amazon's. An agent loop that walks the tree with a second judge model is slow. With Jev, each level of the tree is a choice question that returns probabilities for the child categories, and a combined path score flags weak paths. Products with a low path score go to a person. The author found the agent still better at choosing between two plausible branches, so a merchandiser checks low-scoring and high-value items before the category is saved.
Example workflow
- Export product titles and descriptions from Shopify or the catalog
- Ask Jev to choose the best child category at each level of the taxonomy
- Compute a path score from the chosen probabilities
- Auto-save high-scoring paths; queue low scores for review
- Merchandiser checks queued items and spot-checks saved ones
Tools
- Jev (TypeSafe AI)
- Shopify product taxonomy
- Amazon browse tree
- Python
Risks and guardrails
- Close calls between sibling categories can go the wrong way
- A path score can approve wrong paths too readily
- Taxonomy changes require re-running the tree
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- Blog post by Sam Reghenzi comparing an agent loop and Jev: blog.r6i.it/typesafe-jev-vs-agentic-loop.html
- Hacker News submission: news.ycombinator.com/item