TinyML (on-device machine learning)
also: TinyML · edge AI · embedded ML
Running small machine-learning models on microcontrollers and low-power chips for tasks like keyword spotting or anomaly detection, without the cloud.
TinyML brings inference to devices with very little memory and power budgets measured in milliwatts. Typical uses: detecting a wake word, recognizing gestures from an accelerometer, classifying machine sounds or vibration, spotting anomalies in sensor data, or simple vision tasks. Processing locally saves radio power and bandwidth, protects privacy and works offline.
Development means collecting and labeling representative data from real devices, training a compact model, quantizing it, and validating it on the target hardware. Some newer microcontrollers include neural-network accelerators.
Where it bites: data collection is usually the largest effort, models trained on lab data often fail on field data, and updating models requires an OTA path.
Founder takeaway
Collect real-world data from your actual sensors and mounting early; it decides whether on-device ML will work more than the choice of model does.
sources
engineer review pending