// a.i. on the mcu
We Train Models for the Microcontroller
Most "edge AI" stops at running someone else's model. We own the whole loop:
collect data from your sensors, train a neural network, quantize it to int8 with
Espressif's esp-ppq, and ship it as a .espdl binary
running on ESP-DL — inference happens on the ESP32-S3 itself, no cloud, no latency,
no per-call cost.
01 · train
PyTorch / Keras
Network trained on your real sensor data — behavior learned from examples, not hard-coded rules.
02 · export
ONNX
Framework-neutral export, ready for the Espressif toolchain.
03 · quantize
esp-ppq · int8
Weights compressed to int8 with Espressif's PPQ fork — sized for kilobytes of RAM, not gigabytes.
04 · deploy
ESP-DL on ESP32-S3
The .espdl model lives in flash and runs on Espressif's inference runtime — on the same chip that reads the sensors.
Sensors in, decisions out
Example from our lab: a network that reads temperature, humidity and light, and decides —
on the chip — whether to switch the fan, the pump or the lights. The rules are learned
from data. Swap the sensors and the dataset, and the same pipeline fits your product —
including vision (person and object detection with the ESP32-S3 camera, via ESP-DL)
and audio (keyword spotting).
vision
keyword spotting
predictive maintenance
anomaly detection
classification
python train_export.py
training sensor_model... done
export → sensor_model.onnx
esp-ppq int8 → sensor_model.espdl (target: esp32s3)
idf.py flash monitor
I (1234) minibrain: T=37.2C U=55% L=640lux -> CMD_FAN_ON
I (3234) minibrain: T=22.1C U=18% L=400lux -> CMD_PUMP_ON
~ #
