E-Store Agent

What it does

ListingGuard processes 1688 pet-toy product data for Amazon US preparation. It:

• organizes and deduplicates supplier product and SKU information;
• separates raw source files from edited and final media;
• drafts English titles, highlights, bullets, descriptions, attributes, and SKU mappings;
• checks image format, dimensions, white-background main-image rules, text/logo/watermark risks, and product-image consistency;
• checks missing or conflicting material, size, weight, package, brand, IP, and variation information;
• generates a local detail-page preview and machine-readable compliance report;
• assigns a publication gate, including BLOCKED, so uncertain products cannot silently proceed to upload.

How we built it

The prototype uses a file-based, repeatable pipeline. Source product records remain read-only, while generated listing drafts, compliance prechecks, image drafts, and HTML previews are stored as separate artifacts. Stable internal product codes preserve the link from Amazon listing drafts back to supplier evidence.

The pilot covers three deliberately different cases: a TPR squeaky bone toy, a plush squeaky animal toy with single-item and bundle variants, and a sliding-compartment puzzle feeder with conflicting dimensions and structure mappings. The validation scripts generate the pilot assets and render the review previews.

Challenges we ran into

The hardest problem was not generating text or images; it was preventing plausible-looking but unsupported output. Supplier pages combine marketing copy, multiple variants, bundle images, and incomplete attributes. AI image cleanup can also change mechanical structures or invent details. We therefore added explicit evidence status, conflict checks, variant separation, and a blocking gate.

Accomplishments that we're proud of

• We made missing evidence visible instead of hiding it behind polished listing copy.
• We separated a true Amazon main-image candidate from supplier marketing images.
• We identified that single items, bundles, and materially different structures must not share one variation family.
• We used the mechanically complex puzzle feeder as a negative test: the correct result is an automatic block, not a confident-looking listing.
• We produced repeatable JSON reports and local HTML previews that a human reviewer can inspect before publication.

What we learned

Supplier text is not automatically verified product fact. Image generation is useful for cleanup and layout drafts, but mechanical products need stricter visual comparison and human review. A reliable commerce workflow needs provenance, uncertainty states, and hard stops in addition to automation.

What's next

Next we will complete supplier evidence and sample checks for the pilot products, finish commercial image-rights confirmation, connect the approved output to Amazon upload templates or the Selling Partner API, and record upload errors for feedback. We will then expand from the three-product pilot to a larger review queue while keeping brand/IP, safety, dangerous-goods, and unsupported-claim cases behind mandatory human approval.

Built with

Python, JavaScript/Node.js, JSON, HTML, SQLite, Playwright, image-processing and OCR utilities

Built With

• codex