The challenge
Media Track digitises printed newspapers and runs media-monitoring services. Two steps held it back. Building a media-monitoring profile meant hand-crafting technical boolean search queries, which was error-prone. And its newspaper digitisation pipeline still needed a person to review most pages, because the model that tags and segments articles was not accurate enough to run on its own.
What we did
- Built a Gemini assistant, embedded in Media Track's existing interface, that guides the user through a structured set of questions and turns plain-language intent into a correct, deduplicated boolean query, with a review-and-approve step before it runs
- Persisted each step to Firestore so the conversation state is reliable, and tuned the model to lift the precision of results
- Fine-tuned a self-hosted open model, Gemma, to tag page blocks and segment articles, including articles that run across two facing pages
- Added a composite confidence score that routes only low-confidence pages to a human, so more pages process automatically, served on an autoscaling GPU fleet inside Media Track's own Google Cloud
The results
- A production search assistant that makes media-monitoring profiles usable by non-technical staff
- In a proof of concept, the search approach reached recall of around 97 percent, with the production work focused on lifting precision
- A self-hosted digitisation model that raises the share of newspaper pages processed without manual review, with full model ownership kept in-house