Executive Summary

V1

SIX

A proof-of-concept generative commerce platform that solves the "Netflix Paradox" by dynamically generating personalized shopping interfaces based on user intent.

V1 Architecture: Manual API calls with one-shot layout generation. See V2 Documentation for the Agent-Native architecture with CopilotKit and real-time refinement.

2
Personas
34
Products
<3s
Generation Time
~$0.01
Per Layout

The Problem

Traditional e-commerce is static and "data-starved."

  • Data Density Gap: Retailers rely on sparse purchase history (one coat every 3 years), while streaming services use dense behavioral signals (skipping songs, watching episodes).
  • Inventory Physics: Unlike digital media (infinite supply), physical inventory is finite. Personalization fails if it recommends out-of-stock items.
  • Intent Mismatch: Users alternate between "Hunter" (transactional, urgent) and "Gatherer" (discovery) modes. Static pages cannot accommodate both.

The V1 Solution

Move from Static Pages to Generative UI.

Instead of fetching a pre-built template, the frontend sends user context to an API endpoint which calls GPT-4o to "merchandise" the layout in real-time. The store shelves rearrange themselves the moment the user enters.

  • Intent-Driven: User explicitly selects persona (Hunter/Gatherer), solving the data density problem.
  • Inventory-Grounded: AI only selects from in-stock products, solving the physics constraint.
  • Context-Aware: Weather, time, and urgency influence product selection.

Two Distinct Experiences

The Hunter

"I Know What I Need"

  • • Dense 4-column grid layout
  • • No hero images—efficiency first
  • • Quick-buy cards with Add to Cart
  • • Products sorted by stock & relevance
  • • Urgency banners for context
Try V1 Hunter Experience

The Gatherer

"Inspire Me"

  • • Hero section with mood-setting headline
  • • Magazine-style editorial layouts
  • • Curated sections with storytelling
  • • Aspirational product photography
  • • Discovery-focused browsing
Try V1 Gatherer Experience

V1 Architecture Decisions

DecisionChoiceRationale
AI ModelGPT-4oBest structured output support for reliable JSON generation
Generation MethodgenerateObject() with ZodType-safe schema validation prevents hallucinations
API PatternManual fetch() callsSimple, one-shot request/response
InventoryJSON file (not database)Zero latency for POC; easy migration path to Supabase
Hallucination Prevention3-layer validationPrompt + Schema + Post-validation with retries

V1 Limitations

What V1 cannot do (addressed in V2):

  • No Refinement: User cannot say "too expensive" and get a modified layout
  • Static Inventory: Stock levels don't update in real-time
  • One-Shot Generation: Each request is independent—no conversation memory
  • Manual Orchestration: Frontend manages all state and API calls

See V2 Documentation for Agent-Native architecture with CopilotKit

Continue Reading