Logistics · Global operations
No two logistics emails look alike, and rules could not keep up
Fr. Meyer’s Sohn processes thousands of operational emails carrying shipment, routing and scheduling data, in German and English, in no fixed format. Rule-based extraction could not handle the variety. We built a GPT-powered extraction pipeline deployed on their own servers.
Key results
Client: Fr. Meyer’s Sohn
An international logistics and shipping company with global operations.
- Industry
- Logistics and shipping
- Use case
- Email data extraction
- AI approach
- GPT-3.5 and GPT-4 with prompt engineering
- Deployment
- Docker on the client’s servers
- Languages
- German and English
- Engagement
- Proof of concept into production build
In short
Why fixed templates were never going to work
- The emails carry shipment details, routing information and scheduling data, and they arrive completely unstructured, in German and in English.
- No two look the same. Format varies by sender, country and language, not every required field is present, and date, address and reference conventions differ per market.
- Anything built on fixed templates or keyword matching fails at the volume Fr. Meyer’s Sohn operates at, which is why the rule-based approach did not survive.
- GPT-3.5 and GPT-4 with prompts tuned for logistics now read the email and return the fields as clean structured data, and a missing field is flagged rather than guessed.
The starting point
The challenge
The company needed specific predefined information extracted from thousands of unstructured, multilingual emails and returned in a structured format. No two emails looked the same, and rule-based extraction could not cope with the variety.
Fr. Meyer’s Sohn processes large volumes of operational email every day. Those messages contain critical logistics data, shipment details, routing information, scheduling data, and they arrive in completely unstructured form, written in both German and English.
The company needed defined fields pulled out and handed to downstream systems in a structured format. The difficulty was the variety: format varies by sender, country and language, not all required fields are always present, and local standards for dates, addresses and reference numbers differ across markets.
Traditional rule-based extraction could not handle that. Anything relying on fixed templates or keyword matching was going to fail at the scale they operate at.
The build
What we built
We started with a proof of concept to show that generative AI could handle the extraction reliably. Once the client had validated the results, we moved to a full production build.
Extraction tuned for logistics language
The system uses GPT-3.5 and GPT-4 with prompts engineered for logistics-specific data. It reads the email text, finds the data points the client defined, and returns them in a clean structured format.
Two languages, and no guessing
German and English content are processed natively, including the differences in formatting, date conventions and terminology that come with global operations. Where a field is missing it is flagged rather than invented, which is what keeps the downstream data trustworthy.
On their own servers, behind an API
The solution is built with FastAPI and packaged in a Docker container for deployment on the client’s own servers. Fr. Meyer’s Sohn sends a request and receives structured data back, inside their existing operational workflow.
Built to run unattended
Logging, error handling and maintenance tooling were part of the build from the start, so the system runs reliably at scale without someone watching it.
What changed
The results
Before
Manual extraction from unstructured emails. Time-consuming, error-prone, and impossible to scale as the email volume grew.
After
Automated extraction through an API. Structured data in real time, both languages handled, running on the client’s own infrastructure.
The system identifies the logistics-relevant information regardless of language or format, and manual extraction effort fell dramatically.
The structured output feeds directly into Fr. Meyer’s Sohn’s operational processes, so nothing has to be retyped between the email and the system that uses it.
The architecture is flexible enough that the client can adapt it to new data requirements or additional languages as their operations change, without a rebuild.
Questions about this project
Email data extraction. The company needed specific predefined information extracted from thousands of unstructured, multilingual emails and returned in a structured format. No two emails looked the same, and rule-based extraction could not cope with the variety.
German and English processed automatically: 2 languages. Less manual extraction effort: –80%. From proof of concept to production deployment: PoC → prod. Running on the client’s own servers in Docker: On-premise.
GPT-3.5 and GPT-4 with prompt engineering. Technology used: GPT-3.5, GPT-4, Prompt engineering, FastAPI, Docker, Python, On-premise deployment.
Technology used
Talk to the people who built this
theBlue.ai comes back within one business day.
