Artificial Intelligence / Architectures / Edge
Edge AI: Artificial Intelligence Where Data Is Generated
Factory, warehouse, store, construction site, remote location: The decision is made on-site, with or without a communication link.A camera on a production line, a sensor on a machine, or a scanner in a warehouse generates data that no one has time to send to a web datacenter —and then wait for a response. The data is processed right at the source, in a matter of milliseconds, and continues even when the Internet goes down. We design the equipment, its remote management, and its integration with your production systems and information systems.
There are two reasons why calculations must be done on-site, and a third reason makes this option less expensive
Edge AI isn’t just a trend: it’s what remains when latency, connectivity, or volume make a central solution impossible. When none of these three factors apply, we don’t offer it.
Latency: The decision must be made before the next station
A part passes in front of the camera, the line moves forward, and sorting must be completed within the conveyor’s cycle time. A round trip to a datacenter —even in Switzerland—takes hundreds of milliseconds, and the time varies. The model must operate right next to the camera; otherwise, it’s useless.
The line doesn't wait for a response.
Connectivity: Connection is not guaranteed
A construction site, a mountain facility, an energy plant, a ship, or a factory whose connection is shared with everything else: the connection goes down, or it was never designed for real-time operation. On-site processing continues; the results are sent back once the connection is restored.
What matters works offline.
Volume: Sending all the images costs more than analyzing them
An industrial camera generates more data in a single day than an administrative department does in a year. Sending all of that data to headquarters means paying for bandwidth and storage for images that, for the most part, show nothing. We analyze the data on-site and send only what matters.
We trace events backward, not data flows.
Places where this architectural style can be found
Six locations, six different challenges related to hardware, administration, and communication. The one thing they all have in common: there’s no one on site to restart a server.
The Factory and the Production Line
End-of-line inspection, counting, code reading, and anomaly detection on a machine. Cycle time dictates latency; dust and heat dictate the hardware.
The warehouse and the loading dock
Checking a shipment, counting pallets, reading damaged labels, alerting to a congested area. Numerous cameras, continuous video streams, and a connection that’s often overloaded.
The store and the retail location
Shelves, checkout lines, camera-based inventory, staff assistance. Many small sites, each with compact equipment that is managed remotely.
Energy Systems and Technical Infrastructure
Equipment monitoring, drift detection, instrument readings. Remote, regulated facilities where communication is not a given and safety is paramount.
The Construction Site
Site safety, availability of protective equipment, progress tracking via photos. A site that changes every week, a mobile connection, and equipment that’s constantly being moved.
Installation without a reliable connection
In the mountains, on a ship, or at a remote location: the system must operate on its own for hours or days, synchronize whenever possible, and update itself without anyone having to visit it.
Use cases that support it
These projects take shape as they progress, and we present them for what they are: our role is to provide the platform, integration with the information system, and on-site infrastructure. For optics and computer vision algorithms, the project brings in a machine vision specialist when the project requires it.
Camera-Based Quality Control
Once a defective part is detected on the line—whether it is blocked or flagged—the incident is reported toERP and the manager at Teams.
Equipment Monitoring and Anomaly Detection
On-site sensors and preprocessing, historical data and models at the center, alerts before a shutdown—emphasizing “monitoring” over “prediction.”
Industrial Sites Without Internet Access
A task, a literature search, or a check that continues when the connection is lost and syncs up when it's restored.
Intelligent Document Processing
Scan delivery slips and labels at the loading dock, on-site, when the connection doesn't allow time to wait for the center.
Published Project · 2020A standalone controller, with or without an Internet connection
Equipment controlled on-site by a small computer, with telemetry sent to Azure IoT Central when a connection is available. No artificial intelligence: the true principle of on-site processing.
The LibraryAll Use Cases
Organized by topic: observing the physical world, using artificial intelligence locally, and other categories.
Working Offline: What It Means, Word for Word
"No Internet" covers two very different situations, and publishers' documentation does not treat them the same way. We distinguish between them before making any promises.
Intermittent connection: documented, and this is the typical scenario
The site is connected to the center, but the connection may drop, become overloaded, or be unreliable. Models are cached on-site; inference continues during an outage, and updates and monitoring resume when the connection is restored. Microsoft documents this behavior for Foundry Local at Azure Local: Models continue to be served through system restarts and wide-area network outages.
- What continues: inference, monitoring, and local alerts.
- What the interface will handle: event logging, model updates, and central monitoring.
- What needs to be decided: how long the site can operate on its own, and what happens if it exceeds that limit.
Total isolation: possible, under certain conditions, but not for everyone
No connection, ever. Since February 2026, Microsoft has been offering offline operations for Azure Local, subject to eligibility requirements: a corporate agreement, a documented regulatory or sovereign requirement, and an organization ready to operate independently. Not all AI workloads are supported in this mode, and support may change from quarter to quarter.
- What we are writing: What loads are supported in isolation at the design stage, with supporting references.
- What we don't say: "works entirely offline" without having checked each load individually.
- What we often suggest instead: a minimal connection, dedicated to governance, that makes the standard case applicable.
The capabilities ofAzure Local in offline operations and Foundry Local are documented by Microsoft and re-verified before each proposal. This is one of the fastest-changing areas; a promise made in writing one quarter may no longer be valid the next.
What Edge AI Requires Beyond a Centralized Setup
A server in an air-conditioned room with a technician standing nearby is the easy case. Here, everything that has been set up at the center must be rebuilt.
Equipment Designed for the Environment
Dust, heat, vibrations, unstable power, and no server room or climate-controlled cabinet. Dell PowerEdge XR servers are designed for these conditions: compact, ruggedized, with one or two inference accelerators, and remotely manageable via their built-in controller.
Administration Without On-Site Staff
Restarting, updating a template, changing a setting, reading a log: everything is done from the central location, using Azure Arc to manage dozens of sites as if they were one, or with Dell tools when the connection doesn’t support Arc. On-site visits are the exception that we plan for.
Integration with Production
A decision that doesn’t trigger any action is useless: blocking the part on the PLC, creating the event in theERP or CMMS, notifying the manager in Teams. This is integration work, and it’s been our business for a long time—long before artificial intelligence.
The life cycle of the models, multiplied by the number of sites
A computer vision model is trained at headquarters, validated using images from the site, deployed on each piece of equipment, and replaced when production changes—a new part, new lighting, or a new camera. This cycle is designed from day one, not when the twentieth site comes online.
The Safety of Exposed Equipment
A server in a workshop is physically accessible. Secure boot, encrypted drives, device identification, and a network separate from the production network: security measures are designed for a location where you cannot control who comes and goes.
What Powers Edge AI
Technology comes last. Depending on the location, the connection, and what’s already in place, that’s what we implement.
Dell PowerEdge XR, with NVIDIA accelerators
The product line designed for remote and industrial sites: compact size, rugged construction, extended temperature range, and built-in remote management.
Azure Local, in online or offline operations
The Microsoft platform on validated Dell hardware, managed since Azure when the connection is established, but capable of operating without it under certain conditions.
Foundry Local and the services Azure powered by Arc
On-premises models, and services such as video analysis or document processing, are run in containers right next to the data source.
Azure IoT and Industrial Protocols
Collect data from PLCs and sensors, preprocess it on-site, and report events to the central system. That’s what we did for our published IoT project.
Azure Arc
Manage dozens of sites as if they were one: the same strategies, the same updates, the same logs—all from a central location.
Your production systems and your information system
The PLC, the CMMS, theERP, Dynamics 365, Teams : where decisions made on-site must have an immediate impact.
PowerEdge , and XR product lines, as well as the configurations validated for Azure Local, and the capabilities of Foundry Local are documented by Dell Technologies and Microsoft and re-verified before each proposal.
How We Do It
An on-site project begins with a site visit, not with a quote. And the prototype runs on the production line, not in an office.
The Tour and Use Cases
What needs to be decided on-site, how quickly, what happens when the connection goes down, and what actions the decision should trigger in production and the information system. This should be discussed with the production manager, not just the IT department.
The prototype on the line
Some equipment, a camera or existing sensors, a model, and a few weeks of actual production. We measure accuracy, latency, false positives, and how operators use the system.
Equipment, Integration, Administration
Ruggedized hardware tailored to the prototype, connected to the PLC and the information system, manageable from the control center, and secure enough for an exposed location.
Deployment at other sites and operations
The same equipment, the same procedure, site after site. Central monitoring, model updates, and someone to step in when a site stops responding.
The scope is defined before we begin. This includes what we are responsible for, what you are responsible for, what the vision partner does (if there is one), and the clearly stated boundaries.
Related pages
Related architectures and the hardware that supports them.
At your location: the same principle, in a datacenter
When local processing doesn't require you to be right next to the machine, but in your machine room.
View the page HybridDecisions made on the spot; learning at the center
Sensors and on-site inference; history, training, and supervision at the center.
View the page Dell Servers PowerEdgeA server at your remote locations
The XR series is one of the four, and the sizing line is also suitable for workshop equipment.
View the page Dell NetworkThe network connecting the workshop to the center
Access to the core, and the bandwidth and latency requirements of artificial intelligence.
View the page International ShippingYour overseas sites, prepared before shipment
The equipment configured here, delivered and set up there, including customs duties.
View the page The Comparison ToolWhat kind of architecture is right for your project?
Seven questions, one suggested answer, and the four architectures on the same lines.
View the pageLet's talk about your installation
Tell us what decisions are made on site, and how long it takes
The production line, the machine, or the location; what needs to be detected or decided; the time available; and what the connection can currently handle. We’ll come back with the architecture we’d examine, the equipment we’d test, and what a prototype in an actual production environment would allow us to measure.
What we offer is the opportunity to meet the engineers who will do the work. The site visit, the prototype, and the integration are all part of the project. The discussion that precedes them, however, costs nothing.
Renens, Sion, Châtel-Saint-Denis
Microsoft Solutions Partner and Dell Technologies Gold Partner. Ruggedized equipment, its management, the model, and integration into production—all handled by a single point of contact.
Renens VD +41 21 806 37 15
Sion VS +41 27 552 00 22
Châtel-Saint-Denis FR +41 26 322 59 05