01 / Context
The problem.
Merchant videos contain useful product information, but captions, media formats, and catalog references are inconsistent. Turning that content into a commerce capability requires catalog matching, media delivery, and a recoverable processing workflow.
02 / My contribution
What I built.
I designed and implemented an end-to-end workflow that ingests merchant social content, identifies video candidates, extracts SKUs with an LLM, matches them to the merchant’s existing products, and delivers media through object storage and backend APIs.
03 / Approach
Engineering decisions.
Scope the match to the merchant
A product code is not globally unique. Matching within the merchant’s catalog preserves the context needed to identify the correct product.
Make progress persistent
Persistent processing states, incremental crawling, and structured reports support recovery and make repeated runs inspectable.
Represent real catalog complexity
Multi-product posts and duplicate SKUs need explicit association rules rather than a one-video, one-product assumption.
Connect the complete delivery path
Media ingestion, S3-compatible storage, backend callbacks, scheduling, and reporting are part of the operational system.
04 / Outcome & context
What this work shows.
Historical crawls covered roughly 50,000 posts and 5,000 video candidates. These figures describe processing scope, not the number of successful catalog matches or ingested videos.
The work connects LLM extraction with the integration, data integrity, and recovery decisions required by a real product.
Industry project / The cover is a concept illustration. This case study focuses on my contribution and system-level decisions.