The problem
Matching a very large catalog to several competitor product lists is hard when names alone are unreliable. The same product shows up under different names, sizes and images, and a wrong match is costly. A person still has to confirm the uncertain ones.
What I built
- An image-and-text matching pipeline. It crawls a competitor site, takes in product data, downloads images, finds candidates using image similarity (OpenCLIP with FAISS vector search), scores them, and processes everything in batches. Review screens let a person confirm the final matches.
- A staged, resumable matching workflow for several competitor lists. It normalizes the data and matches by rules (text, brand, size). It sorts rows into found, needs-review and not-found, and uses an AI service in a controlled way for the uncertain ones. It has web and desktop screens, runs that can resume, and safe publishing of results.
How it works
- Crawl
- Product data
- Images
- Image similarity search
- Score candidates
- Batches
- Review screens
- Final decisions
The work is staged, so a run can stop and resume, and each uncertain row lands in front of a person instead of being guessed.
Where it was used
Internal operational work. The final matched workbooks were used by the client or their team.
My role
I designed and operated the workflow and the review process, and used AI-assisted development for much of the code.
Skills shown
Computer vision embeddings (OpenCLIP), vector search (FAISS), Python, Flask review screens, batch pipelines, human-in-the-loop design, controlled use of AI.