Document AI
Read on arrival. Checked before it lands.
Eight trade document types, scanned or digital, in whatever layout your supplier happens to send. Every field is extracted, cross-checked against the records it belongs to, and scored for confidence before anything is written.
What you're looking at
One block of unstructured text on arrival, and seven fields pulled out of it in seconds. Six were confident enough to write straight to the shipment. The seventh, a smudged shipper's reference, was not — so it waits for a person instead of being guessed.
Toggle it yourself, above, or keep reading for how the eight document types and the four-stage pipeline behind it work.
What it reads
Eight document types, each with its own extractor.
Not one generic model pointed at everything. Each document type has a dedicated extractor that knows what belongs on that page and what does not.
Commercial invoice
Line items, unit prices, weights, coils, totals
Packing list
Net and gross weights, coil counts, package detail
Bill of Lading
Vessel, voyage, ports, dates, container and weight detail
Letter of Credit
Terms, tolerances, expiry, latest shipment date
Goods Declaration
HS codes, declared and assessed value, duties and taxes
Contract
Pricing basis, quotation period, premium, quantity
Insurance certificate
Cover, value, policy reference, validity
Financial instrument
Instrument type, value, tenor, issuing bank
How it works
Capture, validate, review, then ask.
Hover or click a stage to see what it actually produces.
Not one model
Three tiers of intelligence. One answer you can trust.
Sending every document to the largest available model is slow, expensive, and no more accurate on a page a parser could have read exactly. TradeDeck tries the most certain route first.
- 01
Direct read
No model involvedMachine-readable documents are parsed field by field, deterministically. Nothing is interpreted and nothing is inferred. The best answer is the one that never needed a model at all.
- 02
Fast vision model
SecondsClean scans go to a fast multimodal model that reads layout as well as text. It handles the bulk of everyday documents at a fraction of the cost and time of a frontier model.
- 03
Frontier model
The hard pagesStamped, skewed, annotated or simply unusual pages escalate to a frontier vision model. The documents that used to be retyped by hand are exactly what this tier exists for.
You never pay a frontier model to read a page a parser could have handled, and never hand a hard page to a parser that could not.
What comes out
Every field on the page, not just the header.
- Quantities, weights and coil counts, per line item
- Unit prices, line amounts and invoice totals
- Vessel, voyage, ports and every date on the page
- Container numbers, sizes and package detail
- HS codes, declared value, assessed value and duty
- Credit terms, tolerances, expiry and latest shipment date
- Delivery terms, origin, grade and quality
- Bank, instrument and reference numbers
The honest test
Bring your worst document. The stamped, skewed one.
Any vendor can read a clean PDF. The useful demo is the page your team currently retypes by hand, so bring that one.