AI · Document automation
DocFlow Processor
Invoice and BOL extraction with human-in-the-loop review — 94% field accuracy, AP cycle time cut by six days.
Meridian Freight · 10 weeks · 2024
Challenge
Accounts payable re-keyed hundreds of freight invoices weekly. OCR templates broke whenever carriers changed PDF layouts, and auditors needed a clear human approval trail.
Approach
Hybrid pipeline: layout detection, specialized extraction prompts per carrier family, and a Next.js review desk where clerks accept or fix fields in one keystroke. Every decision is logged for audit.
Timeline
- 01
Sample & schema design
Week 1–3Collected 800 historical invoices, defined JSON schema and exception taxonomy.
- Schema
- Labeled set
- Accuracy targets
- 02
Pipeline & review UI
Week 4–7Ingestion workers, model ensemble, and clerk review with keyboard shortcuts.
- Extraction API
- Review app
- TMS hooks
- 03
Pilot & hardening
Week 8–10Two AP clerks on pilot lane, SOC2-friendly logging, full carrier rollout.
- Production cutover
- Audit exports
- Training
Stack
AI / OCR
Vision models, Layout analysis, Validation rules
Workers
Python, Celery, S3, PostgreSQL
Review UI
Next.js, React, TanStack Table
Integrations
REST TMS, Email ingest, Webhook alerts
“Nineteen days from kickoff to a working scheduling tool — I didn’t think it was possible. Three months later it’s replaced every spreadsheet in our ops department and I sleep better on weekends.”
Rebecca Holt, Ops Director — Kestrel Manufacturing
Want similar results?
Get a free estimate and timeline within 48 hours.