Edge AI & connected intelligence
Machines, vehicles and sensors report more than anyone can watch. We put AI on the device and at the edge, or in the platform when that fits better, and turn their readings into alerts and answers a person can act on.
What you get
- Device data read at the source, over the protocols the hardware speaks
- Models placed where the data and the connection allow: device, edge or platform
- Alerts and answers that show where each reading came from and how old it is
- Systems that keep working when the network drops
Who it is for
You run physical things: vessels, vehicles, refrigerated shipments, buildings, screens or field equipment. They report data all day, and a dashboard shows the numbers without saying what to do about them. The connection is often unreliable: in a basement, at sea or on a remote site.
What we build
- Integration at the source. Software that reads hardware directly, over the protocols it speaks, with no simulated feed.
- Models in the right place. On the device, at the edge or in the platform, with model size, quantisation and latency engineered against the hardware that exists.
- Questions in plain language. An operator asks what is happening, the system picks the data sources that answer it, and the same state drives the alerts.
- Readings people can trust. Each answer carries the provenance and age of its readings, and a gap in the data is reported as a gap.
- Interfaces where the work happens. A helm display, a TV screen, a phone, a message in the app people already use.
How an engagement runs
| Stage | What happens | Duration |
|---|---|---|
| 01 · Ground the case | Which readings matter, which protocols and devices, where models may run, and a verdict on the data | 3-5 weeks |
| 02 · Prove the value | A prototype tested on your hardware early, to check the constraint is workable | 3-6 weeks |
| 03 · Engineer the system | Integration hardened for one site, vessel or route, and ported to the target hardware | 8-16 weeks |
| 04 · Pilot in one place | Operators work with it in one place, and the metric is measured | 6-12 weeks |
| 05 · Scale what works | Extended to the fleet or network, with device support confirmed | 3-6 months |
| 06 · Run and evolve | Model updates tested before rollout; drift and alert quality reviewed every quarter | Ongoing |
Each stage is contracted separately, and the price is quoted on request once its scope is fixed.
Built for Switzerland
Where data stays is a design decision made in stage 01: on the device, on your network, in a Swiss or EU cloud. On-device inference keeps the data on the hardware, and anyone can check that by confirming no network call was made. For products sold into the EU, the Machinery Regulation and the EU AI Act can both apply, and AI used as a safety component in critical infrastructure needs a counsel check. Dates, articles and sources are on the regulatory reference.
Proof from the lab
| Build | Label | What it shows |
|---|---|---|
| A helm for a fully electric boat | Client delivery | Navigation and charting on OpenCPN and wxWidgets on Linux, reading the vessel over CAN bus and NMEA, live battery, propulsion and position analytics, ported from x86 to ARM. A first working prototype, demonstrated on the water in Spain |
| A cold-chain command centre | Working prototype | Plain-language questions over a temperature-controlled network, excursion detection, predictive stockout and delay signals on a map, with provenance in each answer |
| On-device inference | Working prototype | Speech and language models with no network call at inference time |
| Telebimmer | Live | Car data from the official BMW CarData service with no hardware in the car: fault codes in plain language, alerts for a fuel drop while parked or movement without a trip, geofenced hail warnings, and the age on every value. Not affiliated with BMW AG |
| Lumen | Beta | An ambient display app for LG webOS TVs, fed by an open JSON feed, that plays from a local cache when the network drops. In LG Content Store review |
| Parkberry | Beta | A parking and EV-charging marketplace with map search on PostGIS and a vehicle-fit check for height, length and EV connector type |
| Zeza | Beta | The whole speech, language and voice pipeline running on the iPhone |
Built on
Questions
What have you built on hardware for a client?
A first working prototype of a helm for a fully electric boat: navigation, charts and live battery, propulsion and position analytics, reading the vessel over CAN bus and NMEA and demonstrated on the water in Spain.
Which protocols and devices do you work with?
So far: CAN bus and NMEA on a vessel, a car maker’s official data service, LG webOS TVs, iPhones and Macs. For other buses and devices, stage 01 scopes the integration against your hardware before anything is built.
Does it work without a connection?
It can be designed to. Our on-device prototype runs speech and language models with no network call at inference time, and Lumen keeps playing from a local cache when the network drops.
What happens when a sensor stops reporting?
The system says so. In our cold-chain prototype an answer separates a confirmed temperature excursion from a sensor gap that only looks like one. Telebimmer labels every value with its age.
Which rules apply to AI in a machine?
If you sell machinery into the EU, the Machinery Regulation adds AI safety components to its scope from 20 January 2027, and such a component can also be high-risk under the EU AI Act. The regulatory reference has the detail.
Next step
Find out whether the data can carry it
Sensor data is often the binding risk, so start with the Data Readiness Review: in 1-3 weeks every source is scored on completeness, accuracy, accessibility and lineage. If the use case itself is still open, start with the AI Opportunity Assessment. The price is on request.
See the Data Readiness Review