A vision model reads every completed task card — handwriting, stamps, margin scribbles — and reconciles what the mechanics actually wrote against what RAAS invoiced. Transcription leakage commonly runs 1–5% of labour revenue; this report queues the gaps for your office to accept or reject, and it pays for the whole engagement if it finds 30 unbilled hours a month.
Cards read
8
every card in the batch, not a sample
Items flagged
3
plus 2 low-confidence reads routed to a human
Recoverable this batch
$3,390
2.0 hrs at $120 + $840 part + $2,310 NRC
Annualized at your volume
$122,040
600 cards/mo at an assumed 1.5% hit rate — set your own in Filters
Leakage by type — this batch, CAD
Source: vision extraction vs RAAS invoicing · hours priced at the billed-out rate slider
Review queue — what the cards say vs what RAAS billed
Click a row to open the card in the viewer. Low-confidence reads (<75%) are amber and routed to a human — never auto-accepted. Dispositions are proposals for your office; nothing posts to RAAS.
| Card | Tail | What the card says (vision) | What RAAS billed | Gap $ | Conf. | Disposition |
|---|
Accepted recovery: $0 of $3,390 flagged — queued for office review on this page only; nothing posts to RAAS.
Scanned task card — synthetic facsimile
Invented paperwork, styled like the scan the model would receive. No real tails, licences, or shop data.
What the vision model read
Pre-authored for this synthetic card — not a live model call. Bars are per-field confidence.
What RAAS shows
Source: RAAS invoicing (read-only) — mismatches in red, matches confirmed
How it runs in production
Nightly batch: the day's scanned task cards go through a vision model server-side; extracted fields land with per-field confidence in a reconciliation table in SQL, next to the matching RAAS invoicing rows (read-only). This Power BI report — the same licence already included with RAAS — is the front end. Your office opens the queue with coffee; your AMEs never type anything new.
This page is the honest version of that demo: the eight cards are synthetic facsimiles authored for this pitch, and the "extractions" are pre-computed for exactly these cards. A live model would be theatre on invented paper; the production pipeline is the part being sold, and it runs on your real scans or not at all.
Precision, false positives, and why a review queue still pays
Vision models read legible handwriting at roughly 80–85% field accuracy — which means the queue will contain false alarms. That is priced in: this is a review queue, not an auto-poster. If even 70% of flagged items survive your office's read, the arithmetic holds — one igniter or two unbilled hours a month covers the software's keep, and transcription leakage at 1–5% of labour revenue is the upside case.
Low-confidence fields are never silently dropped and never silently accepted: they route to a human read, amber in the queue, exactly like the coffee-stained card in this batch. Loud uncertainty beats quiet wrongness.
What this is NOT
It never writes to RAAS — read-only on both the scans and the invoicing data. It never gates a maintenance release; the AME's stamp on paper stays the record of work, and nothing here is airworthiness evidence. It is not an OCR replacement for the AME's judgment — the model proposes what it thinks the card says, and a human disposes, every time.
One more honesty note: this batch is seeded at 3 leakage items in 8 cards so one small demo shows every outcome. Real leakage is rarer per card — that is why the annualized figure on page one is labeled a ceiling, and why the first paid step is running the pipeline on a month of your real cards to measure your actual rate.