Built the ingestion layer for an AI SaaS platform as sole engineer. The pipeline takes messy multi-column PDF reports and outputs structured records: coordinate-based page cropping to drop headers and footers, whitespace normalisation to repair text fragmented across columns, and chunking sized for downstream processing. Extraction returns schema-validated objects, each claim paired with its verbatim source quote and page number, rather than free text, so the output is directly usable downstream. Bounded retry and fallback logic means a failed call degrades gracefully instead of corrupting the dataset.