We present an ankle-mounted insole artifact enabling real-time recognition of seven daily activities. Classification is performed on-chip on a dual-core ESP32-S3 microcontroller without requiring any cloud or smartphone connection. Our proposed program design allows for multi-threading, treating sensing (SensorTask), inference (ModelTask), and wireless output (BLETask) as three periodic tasks within FreeRTOS. The SensorTask acquires six pressure channels and a six-axis IMU sampling at 20 Hz. The ModelTask dequeues the sample and executes a post-training-quantized one-dimensional convolutional network in 54 ms. The BLETask transmits the predicted class on change, keeping the radio duty cycle below 7%. Across eight subjects, the system reaches 92.8% leave-one-subject-out accuracy with a worst-case end-to-end latency of up to 555 ms. The mean current increases by 2.24 mA above a 100 mA sensor baseline, allowing for nearly 5 h of operation on a 500 mAh cell. Stability is underpinned by a zero-loss of sensor frames during two hours of continuous streaming with a minimal task jitter below 2 ms. To our knowledge, this is the first insole that reports a complete timing–energy–accuracy triad for entirely local inference on an ESP32-class microcontroller. The results demonstrate that careful task scheduling, rather than network compression alone, is sufficient for achieving reliable edge intelligence in resource-constrained wearables.