Fanziz

The challenge

Fanziz publishes thousands of articles and short-form videos daily across multiple sports and entertainment categories. Connecting the two formats by hand demanded significant editorial effort and was impossible to maintain at scale. Keyword matching produced inaccurate results, and because many videos carried different titles from their corresponding articles, traditional search was unreliable. The goal: automatically identify the most relevant video for every article, with no manual tagging or predefined rules.

What we built

An AI-powered semantic matching pipeline built on OpenAI embeddings and MongoDB Vector Search. Captions are extracted from every uploaded video and converted into semantic vectors, article content is embedded the same way, and both are stored in MongoDB Vector Search. A vector similarity search then identifies and ranks the most relevant videos for each article, which are displayed alongside it automatically. Because the system understands meaning rather than matching exact keywords, it associates articles with videos accurately even when their wording differs significantly.

How it works

A caption-processing pipeline and background embedding jobs generate vectors as new content lands, so associations are created without human intervention. The AI layer handles embeddings, semantic similarity and cross-content recommendation; the search layer runs vector similarity search with automated ranking on MongoDB; and the processing layer keeps content continuously enriched as new articles and videos are published.

Why it matters

Traditional content association depends on metadata, tags and manual editorial work. By leveraging semantic embeddings, Fanziz automatically understands the relationship between written articles and video captions and surfaces contextually relevant videos without human intervention — a richer content experience, improved video discoverability, and an AI-driven workflow that scales effortlessly with continuous content growth.

Results