AB InBev Invoice PDF Automation
Acting architect on a serverless pipeline that normalized invoice PDFs from 30+ supplier formats into structured data for downstream business workflows.
- AWS Lambda
- AWS Step Functions
- AWS Textract
- AWS SQS
- AWS DynamoDB
- AWS S3
- AWS CloudWatch
- Python
Problem
AB InBev received invoices from 30+ suppliers in highly variable PDF formats — different layouts, numeric conventions, partially handwritten fields, and inconsistent structure. Downstream business systems needed clean, normalized data. There was no automated path from raw PDFs to structured records.
My Role
I was the implementation lead and acting architect in practice. I took the initial requirement — “extract and normalize invoice data” — and drove it from requirements clarification through architecture decisions, PDF readability improvements, cost-aware design, and testing-oriented delivery.
What I Built
A fully serverless pipeline on AWS:
- Ingest: PDFs land in S3, triggering a Lambda function.
- Extraction: AWS Textract handles OCR and structured field extraction.
- Orchestration: Step Functions coordinate the extraction, validation, and retry flow.
- Queue: SQS decouples ingestion from downstream processing for resilience.
- Storage: DynamoDB holds normalized records; S3 retains originals.
- Observability: CloudWatch alarms and logs for operational visibility.
I also ran a benchmark phase to quantify extraction accuracy across provider types before committing to the architecture, and designed preprocessing steps to improve PDF readability for challenging formats.
Outcome
The pipeline handled invoices from roughly 30–40 supplier types. The architecture was cost-aware (Textract is priced per page — batch sizing and preprocessing reduced unnecessary spend). Ambiguous formats were flagged for human review rather than silently dropped.