AI commerce
Krafted: an AI product page generator for Shopify
I helped turn a messy, multi-platform product import process into a resumable generation pipeline that could publish complete Shopify product pages without asking merchants to rebuild the result.
The system I joined
Importing a product into Shopify is quick. Turning unreliable marketplace data into a page that is accurate, on-brand, accessible, and ready to publish is not.
The product had to coordinate AI classification, image evaluation, colour extraction, Shopify theme writes, billing gates, and third-party platform limits. A failure halfway through could not force a merchant to start over.
Where I contributed
I worked across the generation architecture, AI decision boundaries, Shopify integration, and reliability controls.
The central design choice was to use models only for subjective work such as classification and copy generation. Deterministic code remained responsible for measurable facts such as contrast ratios, asset availability, billing state, and whether a Shopify write actually persisted.
Product screens
Krafted in Shopify
The embedded application, page-generation flow, store controls, and final product-page output.
Decision trail
Constraints became architecture.
Each decision below connects a concrete limitation to the engineering response and its practical effect.
- 01
Constraint
One generation request crossed AI services, external marketplaces, billing checks, and eventually consistent Shopify APIs.
Decision
Model every stage as an idempotent step with explicit status, retry behaviour, and a durable checkpoint.
Impact
Interrupted jobs resume from the last verified step instead of repeating paid model calls or duplicating Shopify writes.
- 02
Constraint
Language models produced visually plausible colour palettes that still failed accessibility checks.
Decision
Constrain colour generation to validated families, then calculate and correct contrast in code.
Impact
Generated pages meet the 4.5:1 WCAG AA text contrast requirement without relying on prompt compliance.
- 03
Constraint
Shopify theme writes can return a successful response while silently dropping an app block.
Decision
Read every critical write back from Shopify before marking the step complete.
Impact
The pipeline treats persistence, not a 200 response, as the definition of success.
- 04
Constraint
Marketplace titles were keyword-stuffed and regularly pushed classification toward the wrong niche.
Decision
Add a title-normalisation stage and benchmark models against deliberately ambiguous products.
Impact
The classification decision uses both title and description, giving every downstream layout and copy step a stronger starting signal.
Outcome
What the work produced
- A full product-page generation flow across seven niche-specific theme systems.
- A billing-aware publish gate that preserves completed work when a plan limit is reached.
- Verified theme installation and app-block writes despite Shopify's eventual consistency.
- An architecture ready to add more marketplaces and niches through the same normalisation layer.
Technology
The useful lesson was not which model performed best. It was that AI systems become dependable when subjective decisions are bounded and every objective claim is verified by code.
Abrar Ahmed