6 boards ยท 4 kinds of signal
Machine learning on a microcontroller. Today, from your phone.
Record examples with your phone, train a model on our GPUs, and download plain C code for your board. Collecting and training happen in your browser, with nothing to install.
- Boards
- 6
- Signals
- 3
- Bench check
- 32/32 windows
- Output
- plain C
How it works
Three steps, from the first recording to code on your board.
-
01
Collect with your phone
Open a project and scan its QR code. Your phone records movement or sound in the browser, with no app to install, and every recording arrives labelled. Readings from a data logger come in as a CSV file instead.
What you can record -
02
Train on our GPUs
Pick a recipe, or describe your task in a sentence and get one suggested. The report tells you which labels the model mixes up and what to record next, and it refuses to quote an accuracy your data cannot support.
How GPU time is counted -
03
Flash your board
Download a package of plain C: the model and the kernels it needs, with no interpreter and no heap. Your own code computes the features from the sensor and calls the model; you build it with the tools you already use for your board. For the Nano 33 BLE and the ESP32-S3 the package includes an Arduino sketch.
Check your board
What you can count on
Nothing to install to start
Your phone is the data logger.
Collecting, labelling and training all run in the browser. Scan a QR code and your phone records movement or sound straight into your project.
Checked on the real part
32 of 32, on all six boards.
On 1 October 2026 we generated a reference model for each of the six boards and ran it on the board itself. Every board gave the same answer as our server on all 32 test windows.
It tells you what to do next
A report that points somewhere.
If two labels get mixed up, the training report names them, says how often, and suggests recording more of each. If one label has far fewer recordings, it says that too.
Six boards, each checked on the real part
We support a few boards and test each one. Pick yours to see what it can sense, what it cannot, and what it needs on your computer.
-
Tier 1
Arduino Nano 33 BLE Sense Rev2
Cortex-M4F @64 MHz (nRF52840)
Can sense:
- Motion
- Sound
- Sensor values
-
Tier 1
Raspberry Pi Pico 2 W
Cortex-M33 dual @150 MHz (RP2350)
Can sense:
- Motion
- Sound (limited)
- Sensor values
Hold BOOTSEL first time
-
Tier 2
Nordic nRF52840 DK
Cortex-M4F @64 MHz
Can sense:
- Motion
- Sound (limited)
- Sensor values
Needs admin setup
Three kinds of signal
Each page says what you need to record, what the limits are, and which boards can do it.
-
Motion
Gestures and movement, from an accelerometer or your phone's motion sensors.
-
Sound
Short sounds and spoken keywords, from a microphone.
-
Sensor values
Any slowly changing signal: current, temperature, pressure, vibration.
Try it with the board on your desk
The Free plan includes 30 GPU-minutes of training a month, and we estimate a first project at about 8. Check your board first: its page lists what it cannot do as well as what it can.