Case study · Project 04
Clearpost · Invoice automation
High-ROI document ops for messy real PDFs: intake with confidence, side-by-side review against the source file, and output modes for different roles—not another chatbot that “summarizes invoices.”
Problem
Invoice intake is classic automation: repetitive, error-prone, and expensive when humans retype line items. Pure model extraction without a review surface next to the PDF is how bad totals and wrong vendors ship into accounting. Ops needs confidence scores, structured fields, and a path to fix mistakes.
Approach
- File log — every upload lands as a committed PDF with type and confidence so operators see what is ready vs what needs eyes.
- Side-by-side review — source PDF viewer beside vendor, invoice number, date, totals, and line items; validation status and edit-in-review.
- Output modes — technical summary, readable summary, structured records, full review: same pipeline, different consumers.
- Source of truth — never ask people to “trust the bubble”; keep the document open next to the extraction.
Why this design
Boring automation often beats flashy chat. The ROI is hours saved and fewer typos—not a demo of fluency. That fits a quality bar of data quality, human-in-the-loop, and GPU/compute spent where it matters (extraction + review), not endless free-form generation.
Stack
Python · PDF / OCR · document AI · review UI · export paths
Role on the build
Pipeline and product framing (confidence, review, modes). Implementation steered with AI coding agents. ML-aware view of extraction limits and why human review stays in the loop.