- Client
- Global pharmaceutical manufacturer
- Domain
- Document processing
- Method
- Applied AI
34,000 hours of document work, marked for deletion.
Turning hundreds of pages of inconsistent pharmaceutical manufacturing documents into structured, review-ready data.
01
The problem
The problem wasn’t a lack of data.
It was buried in PDFs.
Manufacturing and quality teams were working across long, inconsistent documents from more than 100 external manufacturing partners and markets. Batch records, certificates of analysis, scanned pages, handwritten fields, tables, checkboxes, signatures, and site-specific terminology all had to be interpreted before the information could be reviewed or used downstream.
Some records ran hundreds of pages.
And because the content supported regulated manufacturing processes, “the AI found something that looks right” wasn’t good enough.
02
The human loop
The existing process depended heavily on people to:
openreadlocateextractnormalizeverifyenterrepeat
Across large document volumes, that added up to a projected 34,000+ hours of manual processing every year.
The goal wasn’t to build a chatbot for PDFs.
It was to remove as much of that loop as possible without removing the controls that actually mattered.
03
What we built
A template-agnostic document intelligence pipeline capable of handling digital, scanned, and hybrid manufacturing documents.
Documents enter an event-driven AWS workflow where pages can be processed in parallel using OCR and multimodal vision models.
The system:
- Ingests
- complex multi-page documents
- Extracts
- text, tables, handwriting, checkboxes, and key parameters
- Normalizes
- information from different document formats into consistent schemas
- Preserves
- page-level source evidence and provenance
- Scores and flags
- uncertain values
- Routes
- exceptions and regulated decisions to human reviewers
- Makes
- normalized information available to downstream systems, search, and analytics
Instead of forcing every manufacturer into one document template, the system was designed around the reality that the documents would remain messy.
04
What changed
34,000+
hrs/year
- of manual document processing targeted for elimination
106+
manufacturing partners
- across a highly variable document landscape
100+
markets
- supported across global manufacturing operations
160+
data points
- extractable from complex manufacturing records
The system reduces the amount of information humans need to manually find and transcribe while preserving review where regulatory judgment still matters.
05
Human judgment stayed where it belonged.
This wasn't designed to remove people from regulated decisions.
It was designed to stop making them spend their time finding the information required to make those decisions.
System → extraction, normalization, provenance, confidence
Human → exceptions, validation, regulated judgment
06
Under the hood
- AWS Step Functions
- S3
- SQS
- OCR
- Multimodal LLMs
- Structured extraction
- PostgreSQL
- Vector search
- Human-in-the-loop review
- Auditability
07
The lesson
Automating document work isn't about reading PDFs faster. It's about redesigning everything that happens after someone uploads one.