Talent-profile PDFs, straight into the database
A pipeline that takes actor-profile PDFs in any layout by email, splits them per person, extracts about 20 fields and registers each profile on the platform.
The client’s name is withheld under our confidentiality agreement.
The problem
Talent agencies send profile PDFs in every imaginable layout: one actor per page, one actor across several pages, scanned images with no text layer at all. Staff read each one and typed the details into the platform by hand.
What we built
- An email webhook that acknowledges in under 2 seconds, then processes each attachment in parallel through a queue
- Vision-based page grouping that assembles each person’s profile, even when it spans several pages
- Typed extraction of about 20 fields, including height, languages and birth dates converted from the Japanese era calendar, plus automatic headshot selection
- Per-person PDF splitting and registration through the platform’s API, photos included
Built not to break
When page grouping fails, the batch falls back to one profile per page. Network errors are retried automatically, and anything that still fails lands in a separate queue for a person to check. A regression suite built from real PDFs measures accuracy every time the model or prompt changes.