Automating parcel label extraction at Latvian Post
As the national postal service, the Latvian Posts handles tens of thousands of packages a day, most of which were completed by the sender. However, some slip through the cracks and need data completion in the system, before they can be send out. That was a manual process, up until we shook hands and created an automatic data extraction system, integrated within their workflow.
Last summer, we got in touch with Scandinavian accelerator Future Hub, who was at the time looking for participants in their second wave of the Open Innovation Sustainability program. Together, we discovered that automation and the use of Artificial Intelligence can help tackle a lot of sustainability challenges – challenges their hosts faced on a daily basis.
After a couple of introduction, we met up with Latvijas Pasts – the national Latvian Post, a public company that processes several ten thousand parcels a day and is very well known for the innovative approach towards the future. They, in fact, were one of the first postal services that experimented with delivery robots!
Automating data extraction from incomplete parcels
All the more reason for us to join hands and participate in the accelerator program together, with one united goal: we were to find a solution to complete parcel data, which was often not filled in by sender or receiver and thus couldn’t get processed by the automated system. Instead, employees spent day and night manually going over every parcel, reading and entering the missing information in the management system and moving on to the next package.
Raitis Stūrmanis - Technical Project Manager
To carefully extract missing data from the parcels we process, we currently employ a team of 20 full time employees, who definitely have better things to do.
This manual labour was the focus of our accelerator challenge: do whatever it takes to ease the burden on this team and liberate their time to let them pursue more meaningful work. That meant using image processing of the parcels themselves, automatically extracting data missing in the system and validating the already entered information and forwarding the package status to a carefully managed review cycle, to make sure our algorithm either got it right or learned from its mistakes by doing. All of this was to be accomplished in less than 10 weeks time!
So naturally, we got to work.
As a Microsoft Partner, we fully believe in the power of Azure, in this case: Azure Form Recognizer. Using this building block, we’re instantly able to start (re)training a model that’s already extraction millions of pieces of data from images a year – a well suited solution to the problems. We got to training right away, and took to building the API as well – receiving images and sending back extracted and labeled data was as big of a challenge as the data extraction itself.
Naturally, this all was right up our alley: our data scientists and data engineers have a knack for this, so our solution was testable within weeks. Time enough to enable Latvian Post to send even more image data our way, and decide we had to further improve the speed of the recognition, which we got down to 10 seconds on average, almost 8 times faster than it would take an employee to perform the same task.
(We only achieved this speed within this challenge by using the power of Azure Cloud as Microsoft has the commitment to be as GDPR-compliant as possible.)
Final results of our first collaboration
We’ve got a lot of additional ground to cover, but we managed to already present impressive results at the demo day in December.
Over 85% of the information was automatically extracted, validated and put into the system
liberated hours / month
The bottom line of our challenge: we've got the potential to liberate over 2.000 hours of work each month!
Great! So we presented our demo to the board, the Future Hub team and hundreds of enthusiastic accelerator followers mid December on the Facebook-livestream, on which we were asked what was next.
Easy! We’re looking at:
- deeper integration in the software at Latvian Post (think automatic approval of parcels scanned with a certainty threshold),
- move to an on-premise solution (as most of the critical systems from Latvian Post are),
- increase the accuracy of this proof-of-concept and
- hand-over maintenance to the IT-team from the Post.
A hand-over? For sure; as much as we love to support application we made, we’re even more eager to let you(r team) take over and use our knowledge to further help you automate recurring tasks or predict trends in your data!
No need for waiting any longer: get in touch with us to help you be smarter with time and data on your hands: