Case study · Logistics & freight: An AI document-processing platform spanning AWS and Google Cloud

A multi-cloud platform that extracts data from invoices, bills of lading and customs forms, validates it and posts it to the client's systems, combining 12 AWS services with Google Cloud AI and analytics.

Client: A Canadian logistics provider
Industry: Logistics & freight
Country: Canada
Year: 2024
Timeline: 6 months, ongoing support
Cloud: AWS + Google Cloud

System architecture

How the pieces connect

AWS
Google Cloud
Live flow

01Intake

Amazon SESAWS
Amazon S3AWS
Amazon SQSAWS

02Orchestrate

AWS Step FunctionsAWS
AWS LambdaAWS
Secrets ManagerAWS

03Understand

Cloud RunGCP
Document AIGCP
Vertex AI GeminiGCP

04Validate & review

ECS on FargateAWS
RDS PostgreSQLAWS
Amazon CognitoAWS

05Deliver

Transport system
BigQueryGCP
Looker Studio

(01)Results

What changed.

12 + 4

AWS and Google Cloud services

80%

Of documents processed without manual entry

Hours → minutes

From document received to data in the system

(02)The challenge

The problem.

The client's operations team keyed data from thousands of shipping documents by hand every week: carrier invoices, bills of lading, packing lists and customs forms, arriving by email, upload and scan in dozens of different layouts.

Manual entry was slow, errors caused billing disputes and customs delays, and hiring more staff for peak season was getting harder. Their core applications ran on AWS, while their analytics team already worked in BigQuery, so any solution had to fit both.

(03)The solution

What we built.

We designed a platform that keeps the operational system on AWS and uses Google Cloud for document understanding and analytics. Documents arriving by email, upload or API are stored and queued on AWS, sent to Google Document AI and Gemini on Vertex AI for classification and extraction, then validated against the client's rate cards, shipment records and business rules.

High-confidence results are posted straight into the client's transport management system. Anything uncertain goes to a review screen where an operator sees the document and the extracted fields side by side and corrects them in seconds; those corrections are used to improve extraction over time. Every processed document is streamed to BigQuery so the analytics team can report on volumes, accuracy and supplier costs.

Any document, any channel

Email, upload, scan and API intake with automatic document-type classification.

Validation rules

Extracted data checked against shipments, rate cards and duplicates before it is posted.

Human review queue

Side-by-side review for low-confidence fields, with keyboard-first corrections.

Analytics in BigQuery

Live dashboards on volume, straight-through rate, accuracy and supplier spend.

(04)Architecture

16 cloud services, one system.

AWS

Amazon S3
Document intake and encrypted archive
Amazon SES
Receives documents sent by email
Amazon SQS
Processing queue with retries and dead-letter handling
AWS Step Functions
Orchestrates extraction, validation and posting
AWS Lambda
Validation rules and system integrations
Amazon ECS on Fargate
Review application and internal APIs
Amazon RDS for PostgreSQL
Documents, extracted fields and audit trail
Amazon Cognito
Operator sign-in and roles
Amazon CloudFront
Delivery of the review application
AWS Secrets Manager
Credentials for Google Cloud and partner systems
AWS KMS
Encryption keys for documents and data
Amazon CloudWatch
Metrics, logs and alerts across the pipeline

Google Cloud

Document AI
OCR and form parsing for scanned and digital documents
Vertex AI (Gemini)
Classification and extraction from complex layouts
BigQuery
Analytics warehouse for processing and cost data
Cloud Run
Extraction service running next to the AI models

Stack

Python
FastAPI
TypeScript
React
Terraform
Looker Studio

(05)The outcome

After launch.

Most documents now flow from inbox to transport management system without anyone typing a field. The operations team reviews exceptions instead of keying data, and peak season no longer depends on temporary hires.

Billing disputes fell because every value is validated before it is posted and every correction is traceable. The analytics team gained a live view of supplier costs that previously took weeks to compile.

“The platform reads documents faster and more consistently than we ever could by hand, and our team finally works on exceptions instead of data entry.”
Director of Operations, Canadian logistics provider

Services in this project