Recommendation Engine: Content Matching

Matching job postings to the right candidate websites from a large, constantly shifting pool. The business rules were dense, layered, and full of exceptions. An off-the-shelf solution would have ignored most of the nuance that mattered.

A media company needed to automate a content matching process that was consuming significant manual effort. The challenge was not finding similar content, it was defining what "similar" meant in their specific business context, where a good match depended on criteria that had never been formally documented.

Challenge

The matching criteria were complex and highly contextual. What made a candidate website a "good match" for a given job posting depended on relevance, context, and business-specific factors that varied across categories. These rules lived in the heads of experienced team members, not in any system. The technical challenge was less about building a recommendation engine and more about capturing and encoding domain expertise that had never been written down.

What We Built

We designed a scoring and ranking pipeline using embeddings to measure content similarity, then layered business-specific logic on top to handle the criteria that pure semantic matching would miss. The architecture separates the semantic layer from the business rules engine, allowing domain rules to evolve independently of the model.

Infrastructure took roughly three weeks. The remaining three months went into AI tuning and modeling business rules. The most interesting technical decision was spending the majority of effort not on model configuration, but on translating domain expertise into the scoring system.

What Changed

The system now processes candidate-matching at scale with accuracy the team trusts. Matches that previously required manual review surface automatically, with confidence scores that let the team focus their attention where it matters most. When business rules are complex, most time is spent translating domain expertise into AI behavior, not on infrastructure. That insight shaped every decision we made on this project.