01 / Context
The problem.
Social-media screenshots rarely arrive as isolated, catalog-ready product photographs. The system needs to identify a useful image region and connect it with products in a merchant’s catalog.
02 / My contribution
What I built.
I built an automated visual-search pipeline combining dynamic screenshot-region extraction, image representations, and Elasticsearch dense-vector retrieval. I also worked on descriptive-search embeddings and vector-search efficiency.
03 / Approach
Engineering decisions.
Prepare the visual input
Dynamic cropping separates the relevant screenshot region from surrounding interface and unrelated content.
Retrieve against a real catalog
Image representations connect to Elasticsearch vector retrieval on commerce data.
Integrate the complete workflow
Backend APIs make the matching capability usable inside the product.
Build evaluation separately
A related LLM-based API is in development to classify fully relevant, partially relevant, irrelevant, and no-result retrieval outcomes.
04 / Outcome & context
What this work shows.
The visual-search workflow was implemented for production commerce data, connecting social-media images to catalog retrieval. Related search-quality evaluation and business-opportunity analytics remain in development.
Computer vision becomes useful when input preparation, representations, retrieval, and application behavior work together.
Industry project / The cover is a concept illustration. This case study focuses on my contribution and system-level decisions.