Issuing a series of stamps designed by artificial intelligence is more than a remarkable news story. It illustrates how seriously PostNL takes data science: we don’t just talk about it, we actually do it – experimenting, learning and applying it within the logistics process. Senior Data Scientist Tim Ottens explains how his idea is connected to PostNL’s data ambitions.

Artificial intelligence (AI) is a term popular with companies and their managers. It sounds good and it offers vast opportunities, but initiatives often stall at the level of data analysis and dashboards. At PostNL we actually put AI into practice. That is what makes it such a great place to work for data scientists.

At PostNL, we can spend a part of our time working on innovative projects. The results of this ‘hacking time’ do not necessarily have to be directly aligned to the business objectives. It is mainly about discovering new technologies and possibilities through experimentation. The fact that it can eventually lead to new corporate applications is a bonus.

Works of art

I got the idea for AI stamps when I heard that an artwork designed by a computer had sold for 400,000 dollars. It’s then I realised that stamps are also works of art. So, I got to work with data engineer Ying Yu. In our interaction, she provided the data and software platform, while I mainly focused on the design algorithms. We use the AWS cloud environment for most data activities within PostNL.

Our design model consists of a generating and a discriminating algorithm. Based on existing data, such as images, text, time series or chemical connections, that first algorithm can create something new with somewhat comparable features. The discriminating algorithm establishes whether the artificially generated item is good enough to pass as ‘real’.

That model led to stamps that were devised and approved by a computer. The stamps were all different but homogenous. The difference with a human design process is that we are able to classify things much faster – for example, a portrait, landscape or specific object. We know enough after one or two examples, while an algorithm needs thousands.

Lots of data

We therefore needed a lot of data: thousands of pictures of stamps. All at least one hundred years old that didn’t differ too much from each other but still with enough diversity. Newer stamps differ to such a degree that they are not suitable for our purposes.

We faced challenges that you often encounter as a data scientist: you need enough data of the right quality and enough time to train the generating algorithm. At the same time, the process taught us many new things about visual recognition, which we can certainly use in our sorting centres, for instance.

Expertise department

When our CIO Marcel Krom heard about our experiment he came up with the idea of actually issuing the stamps. Another ambassador was Frank Ferro, our Director Insights – the unit that provides insight into many processes within PostNL. The marketers from the stamp collection department were enthusiastic as well. In the end, we took it on as a standard business project. The stamps will be issued at the end of October, early November. That’s really cool of course!

Indeed, there are more examples from hacking time that will eventually become real business. The ‘phishing classification model’ is extremely interesting, in which suspicious emails sent in the name of PostNL can be automatically assessed as dangerous. Thanks to machine learning, our security specialists no longer have to manually classify whether a mail is phishing or not.

Furthermore, as data specialists we are continuously working on improving the logistics process and making it more sustainable. We are a genuine expertise department working for the entire organisation. As individual data scientists, we also focus on specific domains, such as the website or PostNL app. For example, focused on the ‘next best action’, with which we can optimally respond to the wishes and desires of consumers.

Knowledge sharing

We share a lot of knowledge and experience, including during workshops. Just like with hacking time, it enables you to develop together. For example, concerning standards to be used, new techniques and the value of algorithms in circumstances never seen before, such as a pandemic and lockdown. Together, you discover new implications and any limitations the applied models may have. That is incredibly enjoyable and educational.

Data science and data engineering will become even more important in the future. We can make processes and people increasingly more efficient and effective, enabling us to provide value for PostNL and our customers. That keeps us ahead of our competitors.

Curious? See our job offers.

Do you want to know more?

Check out all our blogs