Edge AI · Product 03 · Vivi engine
Upload your sensor data, train, download C code. Runs on ARM Cortex-M, ESP32 and RISC-V in under 20 KB.
Free to start. No card required.
Your sensor data, as it is.
AutoML picks the right Vivi engine for your data and your chip.
Pure C, zero dependencies. Compile it into your firmware.
No GPU, no operating system. A few KB of flash and RAM is enough.
Learns what normal looks like from your data and flags the rest.
Nodes share what they learn. Offline, no central server.
Vibration, temperature and current together. Under 2 KB extra.
Sounds and movements classified on a chip that costs a few euros.
Every alarm tells you which signal triggered it.
Improves with about fifty new samples, right on the chip.
AI that runs on the device collecting the data, not in a data centre. The decision is made right next to the sensor.
| Cloud AI | Edge AI | |
|---|---|---|
| Connection | Always required | Not needed |
| Latency | 100 ms to several seconds | Under 1 ms |
| Cost | You pay per inference | No recurring cost |
| Data | Leaves the device | Stays where it's produced |
The same engine also runs on servers: it powers Shield and Lynx. The Vivi family →
Motor and pump vibration. You see the fault coming.
Gestures, activity and falls recognised on the wrist.
Vital signs analysed on the device. The data stays there.
Abnormal consumption, occupancy, irrigation. Even with no network.
Motor fault detection, simulated in your browser. Press Play.
No. Upload a CSV, pick your chip and Luviner does the rest: training, optimisation, C code.
ARM Cortex-M0, M3, M4, M7 and M33, ESP32, ESP32-S3 and RISC-V. No GPU, no operating system.
99.6% on the CWRU dataset (bearing faults) and 95% on UCI HAR (human activity), with models under 20 KB.
If your sensors measure vibration, temperature, current, motion or sound and you need real-time decisions, yes.
Through an OTA firmware update, or by retraining on the platform. The model also learns on the chip, from about fifty samples.
A threshold looks at one value at a time. A real fault is often a combination of signals changing over time: the model learns it from your data.