Parent company
Xelec: the reason our AI can leave the data centre.
The AI Experiment is Xelec's AI venture. Software companies stop at the API boundary. We keep going — into the board, the sensor, the enclosure and the production line.
Capability
Model and machine, decided together.
When accuracy, latency, power and cost all have to hold at once, the model and the hardware cannot be designed by two separate companies with a purchase order between them.
What the hardware team does
Four capabilities you cannot buy from a model vendor.
Embedded design
Schematic and PCB design, component selection with real supply-chain constraints, and bring-up of boards intended for production rather than demo.
Sensor and capture integration
Vision, acoustic and vibration capture designed alongside the model that consumes it — resolution, framing, lighting and sampling chosen for accuracy, not convenience.
Thermal, power and compliance validation
Sustained-load thermal testing, power budgeting and the electrical compliance work required to put a device on a European or US industrial site.
Manufacturing and fleet operation
A path from reference design to volume, then OTA updates, remote diagnostics and store-and-forward telemetry for devices in places with no reliable network.
Services you can buy
Hardware engagements, with or without AI.
IoT systems, embedded prototyping, firmware and full electronics product development — run as standalone programmes for clients who need a device built properly.
H-01
IoT systems
A connected product that works on the bench and falls over as a fleet.
H-02
Embedded prototyping
You need proof the thing can exist before anyone will fund it.
H-03
Firmware development
Firmware written once, then nobody can safely change it.
H-04
Electronics product development
A prototype that no contract manufacturer will quote against.
H-05
Sensor and edge AI integration
The model is blamed when the capture was wrong all along.
H-06
Sustaining engineering
A shipped device with no one left who understands it.
Why it matters commercially
Sovereignty stops being a compromise.
Most AI vendors answer a data-residency objection with a region selector. We can answer it with a device: the model runs at the station, the imagery or documents never leave the site, and the audit trail is local.
Regulated environments
Clinics, plants and public-sector sites where off-site processing is simply not permitted.
Bandwidth-constrained sites
Remote facilities, vessels and field operations where uploading raw sensor data is not viable.
Latency-bound processes
Line inspection and safety interlocks with cycle times measured in tens of milliseconds.
Cost at volume
Continuous inference on thousands of streams, where cloud pricing never converges.
Next step
Does your use case need to run on site?
Bring the constraint — cycle time, bandwidth, works council, sovereignty rule — and we will tell you whether hardware is the answer.