On-device AI for retail

Edge0

On-device shopping agents that turn constrained grocery requests into carts ready for confirmation.

Visit live project View repository

Project: Own product
Role: Product strategy, AI systems design, and frontend engineering
Focus: On-device AI, Retail agents, Human confirmation
Stack: Next.js 16, React 19, TypeScript, Neon Postgres, Tailwind CSS 4, Node.js tests

Edge0 on-device AI shopping agents and retail integration toolkit — Product overview and retailer toolkit for configuring bounded shopping-agent behavior.

Edge0 supported shopping tasks on a handheld device — Supported catalog tasks are handled close to the shopper on-device.

Edge0 workflow from retailer data to a cart ready for confirmation — The product flow moves from retailer-approved data to a cart that still requires confirmation.

Edge0 chart comparing specialized small agents with general models — The research story focuses on small, task-specific agents where bounded specialization is useful.

Edge0 internal benchmark comparison — Internal evaluation communicates current demo results without presenting them as production claims.

Edge0 research references and conference publications — Research references connect the product direction to current work in efficient and specialized AI.

The project

Edge0 searches approved retailer product data, applies shopper constraints, and prepares a cart inside the retailer app. Supported work can run on-device, while harder or sensitive requests can move to cloud fallback or a retailer-team handoff.

Runs supported catalog search, product filtering, and cart preparation close to the shopper.
Keeps a confirmation step between an agent's proposal and the final cart action.
Uses bounded on-device execution with explicit cloud fallback and human handoff paths.

Product flow

How it works

01
A shopper describes the products, constraints, and preferences they need.
02
The agent searches retailer-approved catalog data and filters possible matches.
03
It prepares a proposed cart inside the retailer experience.
04
The shopper confirms the action, while unsupported work can use cloud fallback or human handoff.

System design

How it was built

Defined a bounded on-device execution layer for catalog search, filtering, and cart preparation.

Built the public product and deterministic demo in Next.js, with pilot requests validated server-side and stored in Neon Postgres.

Kept confirmation, cloud fallback, human handoff, and benchmark caveats visible so the product does not hide uncertainty.