Redica Systems cut FDA document processing from 2 days to under an hour

The workflow before#

Redica ingests FDA 483s, warning letters, inspection reports and guidance from thousands of sources, much of it scanned PDFs. Commercial OCR read 85 to 90% of the text correctly, so staff corrected extractions by hand. Documents were classified manually or by rules that were right 70 to 80% of the time, and a new filing took 24 to 48 hours to become searchable for customers.

What we built#

We built a document data lake on AWS that takes in every source through Kinesis, Lambda and API Gateway, lands raw files in S3, and retries failed jobs from dead letter queues so nothing is lost.

Each document runs through a Tesseract OCR pipeline trained on regulatory documents, with deskewing and denoising for poor scans. Low-confidence pages go to a person; the rest flow straight through at 99%+ accuracy.

A classification and entity extraction layer on SageMaker tags each document by type, product area, GxP scope, severity and agency, and pulls out companies, sites, investigators and citations. Results land in Redshift, Athena and Elasticsearch, where customers search them in under half a second.

Before and after#

What's next#

With every document classified and its entities linked, the next step is an L3 workflow: an agent that drafts the supplier risk summary a quality team reads before an audit, with a person signing off.