Despite the crisis, sees the growth of AI in the industrial sector

Founded between 2011 and 2012 as part of the Telecom ParisTech incubator, the startup IDMOG moved to Bordeaux in 2014. He specialized in developing software solutions for oil and gas companies. The company has since changed its name to and has grown from 6 to 80 people. is also installed in Paris and Singapore.

“One of my two assistants worked on an industrial site. During the day he processed data manually, and in the evening he developed tools to automate processing,” recalls Aymeric Preveral-Etcheverry, co-founder and CEO of “He realized that this problem that he had, which is the need to bring intelligence to industrial data processing, was found in many companies,” he continues. “From the beginning, we had a desire to develop tools that would enable manufacturers to improve their efficiency by using their data more effectively.”

Since then, has “expanded the scope” of its activities. “We cover various sectors of industrial activity: logistics, supply chain, manufacturing, energy, smart city,” explains Aymeric Preveral-Etcheverry. The company’s clients include large groups such as Suez, Total, the Aéroport de Paris group, RATP, SNCF, as well as medium-sized companies. In total, the company has 50 customers and operates its solutions installed at 150 industrial sites.

“For a client, for example, we are going to answer the problem of reducing the energy consumption on the machine, detecting a compressor failure, turning on the pumps at the right time to reduce energy consumption,” the CEO illustrates.

“Triple experience” combining artificial intelligence and industrial knowledge

But it doesn’t really say what does. “We have triple expertise. First, we understand the needs of the business,” says Aymeric Preveral-Etcheverry. “Secondly, we have data scientists capable of meeting the needs expressed through the development of an appropriate algorithm. Thirdly, our experience in development and DevOps allows us to design an application, deploy it and maintain it in working condition.

In this context, Fieldbox R&D develops software modules, Python libraries, pre-trained models or even templates. “We’re trying to be more than just a consulting firm or a data firm, we’re aiming for a more global technology ambition,” states the CEO. “We contribute to the PyTorch project, publish research papers and share them at conferences.”

In this sense, the company does not develop meta models, unlike Meta or OpenAI. “We try to be pragmatic in order to find the right model for the use case,” says the manager. “When we need to use deep learning algorithms, we look at what is available in the open source ecosystem and we can do transfer learning.”

Recall that transfer learning is about applying the knowledge of a model designed to perform a task—for example, recognizing objects in an image—to another task.

To establish certain use cases, manufacturers sometimes have little data. “In this case, we use standard machine learning models,” says the CEO. is also interested in “domain adaptation”, a sub-technique of transfer learning.

The company advocates additional expertise valued by process engineers. What’s more, the equipment manufacturer Bosch has made it its leitmotif. “Systematically, we have data scientists and engineers on our teams who combine artificial learning models with real physics models.”

According to Aymeric Preveral-Etcheverry, this practice is characterized by a “buzzword”: “physical informative AI.” “There are a lot of fanfare ads on the market, but sometimes you just have to recode a physical line into a model and it does the trick,” he said.

More and more variety of use cases…

Every day, Fieldbox responds to three main project families. “Everything related to the maintenance of equipment, we undertake. First of all, you need to optimize maintenance or even make it predictable if possible,” he says.

The CEO seems to think that the expression “predictive maintenance” is partly overused. “True preventive maintenance is based on supervised learning with a very accurate data set. Theoretically, this is an ideal special case,” notes the CEO. “Before that, we can do unsupervised training to detect weak signal in data clusters where the hardware is not working properly.”

Aymeric Preveral-Etcheverry gives the example of his client Valorem, a photovoltaic panel operator. “With Valorem, we collect all signals from photovoltaic panels, as well as from the electrical equipment surrounding them. Our algorithm allows us to determine which panels are the most faulty in order to optimize maintenance and therefore site performance,” he explains.

The second type of project concerns the optimization of operations, which the CEO summarizes through recommender algorithms. “It could be a recommendation for oven temperature, pump speed, vehicle trajectory.”

“For example, we are working with RATP to develop a model that will help operators decide which train should go to which station and at what time,” he explains. “Obviously there is a schedule, but depending on the hazards – delays, objects on the way – the operator has to react very quickly to redirect the train to the station.”

The considered algorithmic model takes into account all the parameters that the operator checks to make a decision. “That is, how long has the driver been working? Should the train go for maintenance? etc. “.

The model should execute recommendatory scenarios for the operator “very quickly” in order to “reduce its cognitive load”.

“We work with many clients to optimize their energy bills. It’s a big topic for them right now.” Aymeric Preveral-EtcheverryCEO

Moreover, this notion of optimizing operations is gaining popularity among customers due to the current energy crisis. “We work with many clients to optimize their energy bills. It’s a big topic for them at the moment,” the CEO said.

The third topic concerns computer vision. “Industrial facilities have already installed a large fleet of cameras to monitor facilities. This can be used for insidious purposes to automate operations,” he says.

This is what has implemented with Suez. A water specialist inspects small pipes in the network using cameras. The images were viewed only by human operators. “I let you imagine the interest in watching this type of video to be able to indicate that there is a root at kilometer 12 or that there is a collapse at kilometer 45,” says Aymeric Preveral-Etcheverry. “This is something we can fully automate with a computer vision model.” has developed an algorithm capable of detecting objects. It is built into an application that allows the operator to make decisions for “edge cases” that the model has difficulty recognizing. Suez would cut the viewing time by 80%, improve the quality of service as well as the intervention time.

… In production

“POC purgatory has become a reality.” Aymeric Preveral-EtcheverryCEO

The head of the company assures that these are not just POCs or experiments. “The POC purgatory was a reality,” he admits. “Today, I think the situation has changed dramatically. Businesses and operational staff understood the potential benefits. Operations personnel want it now, management must follow: we must find the means to deploy.”

Aymeric Preveral-Etcheverry mentions the fact that about 70% of data processing and artificial intelligence projects among manufacturers fail. “This is where we place our triple experience,” he boasts. “We need to support the business, have functional algorithms, and master deployment.”

Growing influence of the cloud among manufacturers

In addition to more use cases, the CEO is seeing the rise of the cloud among his industrial clients.

“Over the past two years, our customers have been catching up on cloud adoption, especially when it comes to industrial topics,” he notes. “Before the pandemic, we often performed virtualization-based on-premise deployments. During Covid, manufacturers have had time to improve their IT infrastructure. More and more often we are deploying projects independently, often on Azure, but also on GCP, AWS or OVHcloud.”

More or less mature data lakes are used to collect industrial data. “We organize projects in Kubernetes clusters in the customer’s cloud account.”

Edge and hybrid deployments are starting to democratize. “What we’ve done before, which is deploying our on-premises applications on VMware, we’re seeing it come back with solutions like Azure Edge. We often use the K3S project, but we follow the choice of infrastructure by the client.”

In addition, the issue of the sovereign cloud and, more generally, the protection of industrial data is becoming increasingly clear to customers, assures Aymeric Preveral-Etcheverry.

“In general, manufacturers have leveled the technology bricks, now they are looking to get a return on investment.” Aymeric Preveral-Etcheverry

“Several years ago, we made our customers aware of localization and data protection issues. We had more of a prescribing role. Today, the client himself is fully aware of the problems, and makes requests in connection with this. So I guess, I think the offer [de cloud souverain] open joyfully. This corresponds to a real market segment that will grow in importance,” he foresees.

According to the manager, “well-designed” projects require three to four months of development, while more complex themes require about a year of gestation. The company is not afraid of this: predicts a turnover growth of 25% by 2022. Next year, the startup plans to recruit about forty people.

“It’s about responding to market demand: there are different topics related to energy consumption, logistics optimization, etc.,” the manager notes. “In general, manufacturers have leveled the technology bricks, now they are looking to get a return on investment.”

In this regard, software projects are perceived as effective solutions compared to buying new hardware.

Back to top button

Adblock Detected

Please consider supporting us by disabling your ad blocker.