How Everdian delivers “life-saving” real-time critical insights, via AI - interview with Cedric Milinaire

Why does the new generation of European AI startups increasingly turn to Scaleway? It’s not just to access the European cloud’s most powerful GPU cluster. As Cedric Milinaire, Director General & CTO of France’s Everdian explains, it’s also to accelerate growth, as Scaleway’s simplicity means new team members can be onboarded in just a few weeks. Find out more below!

Everdian is an AI startup specialized in real-time decision making. Its main differentiator is that it uses multiple proprietary AI models capable of analyzing large streams of data in real time, to alert strategic decision makers about key ongoing events. Users can build custom dashboards to visualize results and generate their own alerts.

Based on algorithms that could broadly be classified as NLP (Natural Language Processing), its activity covers:

• Real-time security alerts (for example, if there’s a fire alert in a building, the AI helps to get the news out there, and show relevant videos to facilitate fast localization of the fire)
• Multi-source monitoring, to detect fast-spreading topics, and for fact checking
• In the financial sector, predicting market shares and stock evaluations, as well as sentiment detection with regards to big announcements
• Many other use cases, such as finding a client’s stolen assets on second hand reseller platforms.

Everdian uses all types of data, including text, images and videos. For training, the team annotates real world data, then adds synthetic data to improve it. Today, the metadata is often more important than the data itself. So Everdian needs to tweak the datasets to optimize its effects. This can lead to significant improvements in the fields of privacy and energy efficiency.

How it works

AI startups are everywhere right now, as are hype-fueled funding rounds. But Everdian’s objective is to make a difference in the real world.

“When you handle use cases with human lives at stake, ten seconds is really important,” says Milinaire. “For example, we’re used by search and rescue teams to alert them about the occurrence of fire incidents. We provide context with live video feeds and various information posted online. Without us, the only information they may have is that the fire’s in the building. We can tell them - based on data posted online - it’s on the 5th floor and not the 6th. And that saves lives.”

To perform such a feat, Everdian collects data streams into large graphs and analyzes the multiple data points; the level of filtering depends on the services and use cases.

For instance, image analysis services provide more accurate reports than public opinions (often blurry and contradicting). Then feedback correlation and source comparison will provide a clear idea of any situation and enable Everdian to share the most relevant information.

The startup’s proprietary clustering algorithm and AI models analyze image and video similarity, in order to only keep relevant ones. Naturally, the larger the dataset, the harder it is to filter through the noise.

Indeed, the most frequent challenge is understanding the different data points. When Everdian detects critical events, it only wants images of that event, not of people giving their opinion about it. And it needs to select the one best video - not several - that gives the clearest idea of what’s happening. In short, to be able to share only the most relevant and critical information first.

Why they chose Scaleway

Everdian’s number one need is GPUs, “because we analyze millions of texts and images”, says Milinaire, “so we need access to a whole cluster of GPUs in order to optimize our models, syncing them to the hardware. So Scaleway’s H100s are really useful for us.”

They also need highly efficient storage; this is important when handling large amounts of data. For this, Everdian uses Elasticsearch, as it allows for archiving that lets clients “dig through data”, as Milinaire puts it. Everdian uses snapshots on Scaleway Block Storage here.

So the startup’s main pain points were:

1. Cost, as GPUs are expensive
2. Availability: H100 only available as spot instances at first
3. Variety: Everdian’s work requires a wide variety of GPUs (we’re working on it!)

When searching for a cloud provider, Scaleway’s offering and tools largely matched Everdian’s requirements. The main drawback was the security part, as Scaleway was less advanced than other CSPs at that time. Security is a key factor for Everdian, as all new customers demand comprehensive documentation and guarantees on this front.

In the end, the tradeoff was positive, as Everdian’s choice meant they could access advanced cloud features and considerable quantities of GPUs. Individual NVIDIA H100s, as well as entire clusters, are required to analyze millions of texts and images. After that, models are optimized, in sync with the hardware capabilities of each machine.

Building the architecture

To provide a solution able to auto-scale, auto-heal and auto-upgrade, the decision was made to containerize everything and always build on Kubernetes (via Scaleway’s Kapsule product). Then, due to the complexity of data sources, services and customers it has to manage, Everdian opted for a microservices-focused approach.

Their main feature request was for dedicated control planes (in general availability since Autumn 2023) to enable higher levels of resilience and controls. Then, they built everything around those Kubernetes clusters: backups, data and videos, all hosted on Object Storage.

Everdian’s tech teams have notably praised the simplicity and efficiency of Scaleway Kapsule, especially compared with larger CSPs’ equivalent products.

They were also reassured by VPC, where the ability to communicate between different zones, thanks to Scaleway’s Multi-AZ offering, where data is redundant across several availability zones, was perceived as a great advantage. Everdian started in the PAR 1 data center region, then extended to PAR 2 to access those lovely new GPUs, whilst accessing a better level of resilience and reliability.

One missing feature is still the VPN, that Everdian completed themselves for their internal tooling. Their feedback has been noted and Scaleway’s team is working on it.

Milinaire’s current wishlist now includes managed Elasticsearch: a wish Scaleway heard, and so is now looking for others’ points of view in its product discovery approach.

Preparing the next stages of growth by faster onboarding

Everdian found that Scaleway was the ideal cloud provider to ramp up their teams’ technical expertise quickly. “On Scaleway’s platform, our tech teams were operational in a matter of weeks; much faster than with hyperscaler cloud providers,” says Milinaire, who adds:

“We hired a DevSecOps. I didn’t explain anything about Scaleway to him. I just said ‘this is in the Scaleway console, figure it out. You can do it!’ Not long afterwards, he was creating VPCs everywhere!”

Another example was remote employees, who require quick and autonomous onboarding to use other services in a matter of days, without any mentoring or further explanations.

Compared with hyperscalers, this accessibility helps Everdian’s teams be more productive and enables the company to welcome new tech staff more quickly, thereby boosting their impact. With other providers, a non-knowledgeable team member would take weeks to onboard, after reading documentation before being able to start using their first cloud products.

Everdian also cites the proximity of Scaleway’s support staff as a key differentiator: “my feedback is always taken into consideration”, says Milinaire.

This will be critical for Everdian’s next stages of growth, given its ambitious roadmap. Such as reworking the organization and project leveraging new features, along with the always improving IAM and network capabilities of Scaleway.

Another area of improvement will be the AI model optimization - as Everdian grows, their consumption of compute power grows exponentially - needing detailed attention of their AI scientists and technology teams.