Jan De Nul is a company specializing in complex maritime infrastructure works such as offshore energy production, preservation of waterway depths, construction of ports and development of new territories. Its logistics department is constantly busy transporting materials and spare parts internationally, which are subject to customs. One of the tasks resulting from this is that every shipped package needs to receive a specific Harmonised System (HS) code that determines its taxation rate. The attribution of these customs codes is a very complex task: the codes not only depend on the product type, but also on the country the package is imported into. 

‍

Since there are fine nuances between customs codes, choosing the right code for a specific product is complex, and any mistake has an important impact: Mistakes might result in paying a too high tax, a regularization, or a blocked packages at the customs, which is especially problematic when the package needs to arrive quickly.

Because of the complexity of the task, finding the right custom code requires the collaboration of an international customs officer and of local customs agents. 

‍

In order to avoid money and time losses, Jan De Nul was looking for an efficient solution to identify the right customs codes for its shipped goods.

‍

Key element  

• The developed models take into account the cost of an error and choose to automate the decision, or not, depending on the risk of automating the decision

Jan De Nul has a catalog of products with their technical description and internal code. However, the descriptions of the products in the customs catalog are very different from the ones of the company because they do not use the same language register. The goal was therefore to match the customs codes with the internal codes. To this end, the solution developed by Sagacify consisted of 2 steps:

‍

• Retrieve: The retriever was responsible for finding the related context. This acted as a filter and allowed the retrieval of all possible section/subsection candidates. At this step, it doesn't matter if we retrieve too many candidates because we just want to be sure not to miss any. 
• Reader: The reader selected the best candidate. Therefore, it predicted the section/subsection for a given line item from a given context (provided by the retriever).

‍