Air waybills in. Validated, EDI‑ready records out.
AWBGuru reads Master Air Waybills — clean PDFs, skewed phone photos, handwriting — then proves every field against real air-freight logic. Clean documents deliver themselves. You see only the exceptions.
Configure once. Then it runs without you.
Straight-through processing is the default, not the demo. Documents arrive, get classified, split, extracted, validated, and delivered — a human is involved only for genuine exceptions, and never as a blocking step.
Any paper, three doors in
Dump a folder of PDFs, feed the scanner directly, or capture paper in the field with the mobile SDK. Every channel lands in the same pipeline — configure the source once and it keeps flowing on its own.
Nobody labels a document
One PDF often holds several documents. Page-level classification detects boundaries and types, splits the file, and routes each document on — you never tag “this one’s a MAWB.”
One canonical record
An LLM-vision / OCR stack — no custom or fine-tuned model — normalizes every extraction into a single canonical AWB record: parties, routing, weights, charges, line items. Every output format reads from it, so exports never drift.
Freight logic, not guesses
Deterministic, freight-native checks prove each field: pass, warn, or fail. The results feed the document’s confidence score and decide whether it clears on its own.
No export button
The validated record auto-routes to your pre-configured destination — API, EDI message, or spreadsheet drop. There is nothing to click, because clicking is the failure mode.
OCR reads characters. AWBGuru knows freight.
The moat is not the text reading — it’s what happens after. Every field passes deterministic, explainable checks written for air cargo. Each yields pass, warn, or fail, and each feeds the confidence score.
AWB number
PassEleven digits, written 3-8: a 3-digit airline prefix, a 7-digit serial, and a check digit equal to the serial mod 7. The prefix must map to a real carrier. A misread digit fails arithmetic — it cannot slip through.
serial 8134002 mod 7 = 2 ✓ check digit
prefix 020 → LUFTHANSA CARGO ✓ real carrier
Airport & route
WarnOrigin, destination, and routing must be valid IATA codes — and the code must match the printed city. That cross-check catches the “valid-but-wrong” misread no dictionary can: both airports exist, only one is on this waybill. Multi-airport cities (Seoul = ICN/GMP) match at city level.
LAS = LAS VEGAS ≠ LOS ANGELES ⚠ likely misread of LAX
→ routed to review · pipeline keeps moving
HS code
PassThe commodity code must resolve to a valid Harmonized System reference of six or more digits — not just “looks numeric.”
len 6 ≥ 6 ✓
Weight
PassGross vs chargeable consistency, a volumetric recompute from the stated dimensions — (L×W×H)/divisor, divisor configurable per carrier and lane — and line items reconciled to totals.
gross 618.0 → chargeable 760.0 ✓ matches doc
Consolidation
PassIs this MAWB a consolidation of house shipments? AWBGuru decides that first — before attempting single-commodity identification — so a consol manifest is never force-fit into one commodity line.
→ consol = true · commodity ID skipped ✓
Everything feeds the score
Every check result — pass, warn, fail — flows into per-field and document-level confidence. Validation isn’t a report you read; it’s the signal that routes the document.
doc confidence 91.2 → review queue
A calibrated score. No fine-tuned model.
Per-field and document-level confidence, produced without any custom model — and it’s the score that decides what clears unattended.
Validation rules lead
The primary signal is deterministic: did the field pass the freight-native checks? Arithmetic and reference data, not model vibes.
A secondary model signal
Token logprobs or self-consistency blend in as the tiebreaker — useful where no rule applies, never trusted on its own.
Calibrated against ground truth
The blended score is calibrated on a labeled corpus of real waybills, so 98 means 98 — and nothing clears unattended below the proven threshold.
Review is asynchronous
Below-threshold documents queue for a human. The queue never blocks the stream — clean documents keep delivering while exceptions wait their turn.
One canonical record. Every format reads from it.
Extraction lands in a single internal schema — parties, routing, weights, charges, line items. JSON, EDI, and spreadsheets are projections of that one record, so they can’t disagree with each other.
{
"awb_number": "020-81340022",
"origin": "ICN", "destination": "FRA",
"pieces": 4, "gross_weight_kg": 618.0, "chargeable_weight_kg": 760.0,
"is_consolidation": false,
"confidence": { "document": 0.986, "awb_number": 0.999, "route": 0.994 },
"validation": [
{ "check": "awb_mod7", "result": "pass" },
{ "check": "route_city_match", "result": "pass" },
{ "check": "volumetric_recompute", "result": "pass", "divisor": 6000 }
],
"routing": "auto_cleared" // per-field confidence + validation attached to every response
}
XML available on the same endpoint. Payloads shown abridged.
FWB/17 020-81340022ICNFRA/T4K618.0 FLT/LH713/04 RTG/FRALH SHP /SEONG PRECISION EXPORT CO LTD /12 MAGOKJUNGANG-RO/SEOUL/KR CNE /RHEIN COMPONENTS GMBH /HAFENSTRASSE 8/FRANKFURT/DE …
FWB v16 and v17 supported today; the architecture accepts FHL, Cargo-XML (XFWB/XFHL), and ONE Record as they’re added. AWBGuru generates and returns the message — you transmit it through your own rails. We hold no airline credentials.
awb_number,origin,destination,line,pieces,gross_kg,chargeable_kg,hs_code,confidence 020-81340022,ICN,FRA,1,4,618.0,760.0,851762,0.986
CSV and XLSX, one row per line item — drops straight into the workbook your team already runs.
Built to be trusted with other people’s paperwork.
Freight documents are commercial information about your customers. The system is designed to hold them briefly, prove its work, and explain every decision it made.
24-hour auto-delete
Source documents are purged within 24 hours. We keep the structured receipt — not your customers’ paperwork.
Deterministic, explainable checks
Validation is rule-based: arithmetic, reference data, cross-checks. Every verdict shows its work — no model guesses in the audit trail.
Calibrated auto-clear
Nothing clears unattended below the confidence threshold proven on the ground-truth corpus. The gate is measured, not assumed.
Full audit receipt
Every document keeps its record: source, timestamp, extracted fields, and each validation result. Who cleared it — and why — is never a mystery.
No template lock-in
No per-carrier layout setup, ever. A waybill format you’ve never seen processes the same day it arrives.
Filing-ready structure
Records are structured for downstream compliance filings — ICS2 in the EU, ACAS in the US — so advance data starts clean.
Connectors are table stakes.
The canonical record drops into the TMS you already run. Interop is expected — the validation layer is what you can’t get elsewhere.
SMB forwarders first.
Built for the freight forwarder without a data-entry floor — and architected to scale to enterprise volume without changing shape.
The questions forwarders ask first.
Do you transmit the FWB to the airline?
No. AWBGuru generates and returns the message — FWB/16 or FWB/17 — and you transmit it through your own rails. We hold no airline credentials, which keeps your carrier relationships and liability exactly where they are today.
Do we have to set up templates for each carrier’s layout?
No. There is no per-carrier template configuration. New layouts — including skewed phone photos and handwritten fields — are read by the same extraction stack and proven by the same validation rules.
What happens to our documents after processing?
Source files auto-delete within 24 hours. What persists is the structured audit receipt: extracted fields, validation results, timestamps, and delivery record.
What happens when a field fails validation?
The document routes to an asynchronous review queue instead of delivering. It never blocks the stream — clean documents keep clearing while a person resolves the exception, and the correction feeds recalibration.
Which documents are supported today?
Master Air Waybills are the beachhead, with consolidation detection built in. The classify-and-split stage is document-agnostic by design, so coverage expands without re-architecture.
Put your waybills on rails.
Send us a day of your real MAWBs — clean, skewed, handwritten, all of it — and watch what clears without you.