Chapter 11 Filing Extraction Pipeline
🏦 AI/ML Engineering Internship at Bondoro Insights
A document AI pipeline that converts scanned, hundreds-of-page Chapter 11 bankruptcy filings into highly accurate structured text feeding LLM pipelines that generate case summaries and pitchbooks for large-cap bankruptcies. Because downstream summarizers faithfully report whatever the extraction says, the system is engineered around a single principle: a confidently wrong value (a silently corrupted figure or a number drifting away from its label) is worse than a loud failure. It scored 99.3% composite accuracy across a 684-document benchmark.
The architecture is a deterministic OCR spine (RapidOCR) with routed VLM overlays: expensive models run only on pages they provably win, with a MinerU table overlay for dense creditor grids and a GLM-OCR rescue for shredded stylized lines. Custom OpenCV stages recover checkbox states from degraded scans and de-weld superscript markers misread as digits in financial figures, while an independent Tesseract pass cross-checks the spine for silent line-detection misses. A CI-gating evaluation harness enforces exact match on financial fields and label-value association against hand-verified ground truth, and engine choices were settled through structure-aware bake-offs across 5+ OCR/VLM candidates. Persistent worker processes and threaded fan-out for parallel OCR amortize model loads across documents. The pipeline runs deterministically and local-by-default via quantized models on 8GB consumer GPUs, with serverless Modal GPU offload and reproducible parity from Windows/CUDA dev to Apple Silicon/MLX production. (Proprietary codebase — no public repo.)