Community · E2 · artifact verified

Run an e-commerce store on typed decisions alone

Jev Mart is a Japanese e-commerce demo where every operational decision - listings, inquiries, reviews, triage - is a typed Jev call and an LLM is never invoked; it ships with a twelve-chapter lecture set that teaches the pattern.

01 · Role in the system

What Jev does here

The store's operations run on typed answers: incoming inquiries and reviews are triaged by Noul and Choice calls (the API routes normalize answers into kinds with labels and probabilities), product and listing decisions come from the same shape, and an evaluation route scores judgment quality against the ESCI shopping-relevance dataset. The README's headline is the design claim: this app never calls an LLM once - every judgment is Jev returning types and probabilities, which keeps each decision cheap, bounded and auditable. Data is all fictional and labelled as such. Alongside the app sit twelve chapters plus two appendices of lecture material walking through where System One decisions fit in real e-commerce work.

02 · Control boundary

Where Jev sits

Jev as the entire decision layer of a storefront: triage, listing and review judgments are typed calls whose answers normalize into labeled kinds, an ESCI-backed eval route checks judgment quality, and no generative model sits anywhere in the loop.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

  • Demo with entirely fictional data; no real store traffic.
  • Unofficial community sample, not affiliated with TypeSafe AI.
  • Lecture material is Japanese; code comments mixed Japanese/English.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

  • tuneyuki ↗Community · github · public · checked 2026-09-30