Real AI on the smallest chips.
NoxML runs trained models on embedded hardware, so your product can recognize what its sensors are telling it without a round trip to the cloud.
Every trip to the cloud costs time, battery, bandwidth and trust.
A round trip to the cloud adds delay you can’t afford when a motor needs stopping or a fall needs catching. A radio that’s always sending drains the battery. A device that needs the cloud to decide anything stops working when the connection drops. And raw audio or health data leaving the device is a privacy question your customers will ask.
Running the model on the device means the decision happens in milliseconds, right where the data is, and only the answer leaves it.
From notebook to firmware.
Model import
Bring a model trained in your usual tools.
Fits small targets
Quantization and memory planning so the model fits the flash and RAM you have.
Hardware acceleration
Use the DSP or NPU on your chip when it has one.
Sensor pipelines
Filtering and feature extraction for accelerometer, audio and other sensor data.
Model updates over the air
Ship improved models to devices in the field with FleetNox.
Accuracy and latency you can measure
Compare on-device results and inference time to your reference model before you ship.
Devices that know what they’re sensing.
Predictive maintenance
Spot bearing wear or imbalance from vibration before a machine fails.
Wearables and health
Classify activity or detect events on the device, and send only the result over Bluetooth.
Audio event detection
Recognize glass breaking, alarms or keywords without streaming audio off the device.
Specifications
- Licensing
- Contact us
- Operating systems
- NoxOS, Zephyr
Plays well with the rest of the family.
Teach your device something new.
Tell us the sensor, the chip and what you want detected.
