A civil-aviation maintenance (MRO) provider needed a faster, more reliable first-pass check of large maintenance-documentation packages. We built an autonomous computer-vision service — with a CLI and an HTTP API — that classifies each page and flags missing signatures, stamps, empty checklist cells and cross-section mismatches, before a specialist signs off.
First-pass control of MRO documentation packages — work cards, operation sheets, task tables, work orders and acceptance checklists — was manual and slow. A single missing signature or stamp, or one unfilled checklist cell, can hold up a package, and reviewers work through hundreds of scanned pages by hand.
Every page is rendered and classified into its document type — work cards, operation sheets, task tables, work orders, acceptance checklists — with a low-confidence “other” fallback.
Computer-vision detectors locate signatures and stamps on each page and flag where a required one is missing.
Specialised detectors find unfilled checklist cells and verify the “date + signature” pairing on work orders.
Work-card and task numbers are extracted (detector + OCR) and reconciled across the whole package to surface mismatches between sections.
Runs as a CLI for a single file or an HTTP service with a built-in UI; returns a structured JSON report, console output and annotated pages.
This is decision-support: the service produces an advisory report and a specialist gives the final sign-off. It does not perform legal validation.
Delivered as a standalone service (CLI + HTTP API with a built-in UI) that the provider’s own team operates and maintains. It turns hours of manual page-by-page checking into a fast automated first pass, leaving people to focus on the exceptions the system surfaces. Detectors and the page classifier can be retrained as document forms change.
We scope a clear plan with milestones and architecture options — and right-sized GPU hardware if AI workloads are involved.