AI Redact v2: 10× faster, now with custom patterns
Our detection model just got a rewrite. Here's what's new, how we trained it, and why batch jobs now run in seconds.
When we started pdfcraft ai, our goal was simple: make PDFs less awful. Three years in, the product has grown past what we expected, and the team has grown with it.
What changed
The original redaction pipeline was a single pass over each page, detecting entities and masking them in one go. It worked, but performance fell apart on scanned documents above 50 pages.
We rewrote the core detection path from scratch this quarter. The result: 10× faster throughput and a new suite of custom-pattern rules.
The new pipeline
The new architecture splits the work into three stages — OCR, classification, and masking — each horizontally scalable. We also added custom-pattern support so teams can register their own regexes for internal IDs, case numbers, or other domain-specific tokens.
We're just getting started. If you have a workflow you'd like us to support, drop us a line — we read everything.