PDF invoices → governed, dashboard-ready data — with the human-in-the-loop layer that makes it trustworthy.
Real client-facing engagement. I designed and shipped a six-layer Microsoft Power Automate + AI Builder workflow that ingests messy PDF invoices, extracts structured fields, normalises them through a two-worksheet quality layer, flags exceptions for human review, and feeds a live client-facing dashboard.
A six-layer pipeline with quality control baked in.
Trigger and ingest in Power Automate, extract with a custom AI Builder model, preserve the raw output, normalise into a structured layer with formulas, flag exceptions in a tracker, and only then publish to the dashboard. Every layer has a job; no layer pretends to be the layer above it.
PDF → AI extract → structure → QC → dashboard
Six layers, each with one job. Raw AI output is preserved for audit; the structured layer is the only thing the dashboard reads; the tracker decides what humans actually touch.
PDF in, governed data out — with a quality-control layer that decides what humans actually touch.