Customs Bridge bases its AI approach on a French cloud

Customs Bridge is a “LegalTech”. The publisher wants to simplify the understanding of customs regulations for both importers and exporting companies. “AI is at the heart of our economic model, because although they don’t provide an answer at all, algorithms are capable of processing information contained in a large number of documents”, Loïc Poisot summarizes, co-founder and President of Customs Bridge.

The first use case for startup-developed algorithms relates to the classification of products. This task is more complicated than it seems. Products imported into the European Union must be categorized according to a nomenclature with more than 24,000 entries. However, all the complexity of this task for importers is to choose the right category from the description provided by the manufacturer.

In some specific fields such as electronics, these descriptions can be complete and convey all the relevant information to properly categorize the product. In other business sectors, the text can be very short and subject to many interpretations. Loïc Poisot explains the role of AI in this category: “We use AI to rely on rankings generated from training data. We combine this data with more traditional search engines that, with synonymous ontologies, index the information contained in legal texts. “

“It allows us to combine all data sources to provide the best rankings to our customers,” he said. “Nothing replaces human expertise, so the idea is not to provide a solid and specific classification for a product, but to offer a digital assistant that can help people establish customs declarations. »

An economical and sovereign choice

To do this, Customs Bridge joined the OVH program focused on supporting startups. He wanted to gain the privilege of accessing the northerner’s computer resources. Many start-ups opt for Microsoft Azure or AI services over AWS. The co-founders of Customs Bridge also chose OVHcloud because of belief.

“We have decided to ban the use of technologies from GAFA as much as possible and favor French or European technologies.»

Loic PoisotCo-founder and president of Customs Bridge

“We have decided to ban the use of technologies from GAFA as much as possible and favor French or European technologies,” Loïc Poisot said. “This is our business philosophy, but it’s also an argument vis-à-vis our customers”.

Thus, the startup anticipates future uses of its solution in more sensitive contexts.

“Right now, we are only working on customs classification. If in the future we offer a remote declaration tool, companies will put ultra-sensitive export data. Being able to guarantee that this data will be hosted in France is a plus. “, anticipates our interlocutor.

Customs Bridge began to use classic OVHcloud hosting services, then became interested in French CSP’s IA as a Service offer.

The first use case developed by Hamza Saouli, Chief Innovation Officer at Customs Bridge, was about automatic product classification. The expert relies on data accessible to Open Data, particularly the EBTI (European Binding Tariff Information) community database. This dataset contains 250,000 rows covering only 10% to 15% of the complete BOM.

A processing pipeline is set up to cover a customer’s request. It is a question of submitting this request to the model. The answer obtained from the model is prepared and then displayed to the user.

As input, the chain must process files containing the product description provided by its manufacturer. These small files can be processed locally, then uploaded to OVHcloud Object Storage, to be submitted to the model. “In the near future, this pipeline will be more complex, because we’re making a text enhancer,” Hamza Saouli said. “It’s an algorithm that will enrich a data set to increase it from 200,000/300,000 lines to 3-4 million lines through text generation techniques. Training models with such a large amount of data is no longer possible on conventional PCs. »

One of the earliest users of the managed AI offering of OVHcloud

So naturally Hamza Saouli pushed the training of its AI models towards the OVHcloud AI Training offer. “ The transition to learning AI models from an On-Premise approach to OVHcloud AI Training has given us flexibility and power that we could not have in-house ”, he assured.“ The solution is very simple to use : we can set in advance the number of GPUs and the amount of RAM we will need in time t to perform the analysis. This is very useful if you know in advance the amount of resources needed. »

This is how Customs Bridge became one of the very first to test the French CSP’s AI offer. “We entered into a phase of deeper discussions with the OVH team, to bring out the issues associated with this new offer back then, which were often related to our understanding of their platform, but also to guide them towards our needs as ‘users’, Loïc Poisot explains

“The discussions with them were very constructive. “, He added. Among the developments made to the AI ​​Training solution by the OVHcloud teams, is the possibility to train the model on CPUs rather than GPUs, an idea submitted by the startup to limit the cost of model analysis when the data volume is compatible.using this mode.

Customs Bridge was able to explore and take advantage of the capabilities of the AI ​​Training offer as it saw fit. “OVHcloud gives us the option to select the Docker image on which to launch the analysis. I used these containers for the Transformers and Tensorflow models for a Chatbot with images available, and it works. perfect ”, describes Hamza Saouli.

To train his first Transformers models, the expert relied on 2.5 TB of data. Machine Learning models require 30-40 GB of input data.

“With Nvidia V100 GPUs made available by OVHcloud, learning Transformers on 250,000 lines only requires about thirty minutes of calculation. This is very fast for a low cost, as a calculation time is only invoiced at around 1.75 euros. “, he points out.

This very limited cost has allowed the startup to not have to restrict itself to learning its models and experimenting as many times as necessary. “This is the freedom that the cloud model gives us,” concludes Hamza Saouli, who is interested in other OVH As a Service solutions related to AI. He is closely monitoring OVHcloud ML Serving to run production models or even the OVHcloud Data Preparation service that may be relevant for startups, when the volumes of data to be processed upstream of the models will be larger. Dynamic resource allocation will allow Customs Bridge to be charged only for its actual consumption.

“Now, we want to make a real multicloud approach.»

Loic PoisotCo-founder and president of Customs Bridge

If Loïc Poisot wants to remain loyal to OVHcloud in the coming years, the business manager wants to evolve towards a multicloud architecture, to improve the stability of its technical infrastructure.

“At the same time, we have another host for our messaging, the Infomaniak. Now, we want to take a truly multicloud approach. OVH will remain our historic supplier, but we will rely on a second supplier, most likely Infomaniak or Scaleway to set up load balancing and ensure that if there is a problem with one of our cloud providers, we can continue operating for our customer. »

Leave a Comment