Doc Pipeline Kit

Problem

Freight rate-confirmation PDFs arrive in inconsistent per-broker layouts, and turning them into clean structured data (load ID, rates, stops, dates) is normally a manual, error-prone chore. This repo started as a client-acquisition kit for a paid document-extraction service built around that problem — then pivoted the same week.

Approach

The reference implementation is Python 3.12 with PyMuPDF for layout-aware extraction and three broker-layout adapters, checked against a golden-data harness: six fully synthetic PDFs score 100% (157/157 fields), with a fail-loud drift test that refuses to guess an ambiguous dollar amount rather than risk a wrong one. No network calls, no API keys, no LLM in the default path — the whole pipeline runs offline.

Currently building: the Upwork services angle was dropped by owner decision on 2026-07-11, and the repo is pivoting into a free, client-side portfolio tool — a TypeScript/PDF.js port of the same parser, running entirely in-browser so no document ever leaves the visitor's machine. The case-study site is already live on Cloudflare Pages; the browser port itself hasn't started.

Outcome

The repo is public — the parser, goldens, and synthetic fixtures are all on GitHub, and the case-study build is live now; the browser tool is next.