← All engineering posts

4-Part Series

Document AI

Document Processing in Production: Why Every Platform Breaks at 60%
Part 1

Document Processing in Production: Why Every Platform Breaks at 60%

Every vendor demo hits 95% accuracy. Deploy on real documents and it drops to 60%. Here's what breaks, why, and the hybrid pipeline that fixes it.

PDF Table Extraction: Why Structure Recognition Breaks
Part 2

PDF Table Extraction: Why Structure Recognition Breaks

PDF table extraction breaks on merged cells, borderless layouts, and cross-page tables. The detection-then-structure pipeline we built to fix it.

LLM vs OCR Pipeline: Why the Answer Is Both
Part 3

LLM vs OCR Pipeline: Why the Answer Is Both

Multimodal LLMs understand document layout but hallucinate numbers. OCR extracts exact text but misses structure. The hybrid architecture that gets both right.

SEC EDGAR and XBRL: Financial Document AI at Scale
Part 4

SEC EDGAR and XBRL: Financial Document AI at Scale

XBRL promised comparable financial data. Across three companies, only 12% is directly comparable. The NLP mapping layer and document AI pipeline that close the gap.