Project profile
Scope, role and technology
Visual evidence
Duplicate groups, placeholder detections, possible mismatch examples, watermark findings, and audit reports provide evidence for validation before any catalog correction.
Why image quality requires automation
A large e-commerce catalog can contain repeated images, missing variants, unavailable-image placeholders, watermarked assets, or images that belong to a different product. The auditor is implemented and ready, but it has not yet completed enough real-world validation runs to be presented as a production-proven system.
- Separate quality-audit workflow rather than mixing detection directly into publication.
- Inventory Cloudflare R2 assets and compare them against catalog expectations.
- Use deterministic checks first, then visual analysis where semantics are needed.
Audit checks
- Missing expected images.
- Exact or near-duplicate images.
- Placeholder / unavailable-image assets.
- Potential watermarks or undesirable source imagery.
- Potential product/image mismatch and missing visual variants.
Hash-based grouping
Perceptual/content hashes group repeated visual assets even when filenames or URLs differ, reducing the set that requires more expensive or manual inspection.
Human review for ambiguous findings
Similarity and vision results are treated as evidence for review rather than automatic destructive correction. Confirmed corrections can feed back into the Image Updater and catalog.
Architecture
Engineering decisions
- Use cheap deterministic signals before AI vision.
- Treat model confidence as a prioritization signal, not unquestionable truth.
- Keep detection separate from destructive correction.
- Status remains validation-in-progress until enough real catalog runs are completed.