AI Fitness Watch based on RTL8763EW

Client

A sport electronics brand developing a sub-$100 GPS sports watch with AI-powered training insights.

Challenge

Classifying 20+ exercise types (running, cycling, swimming, HIIT, yoga, etc.) automatically using only wrist-based sensor data, with on-device processing for real-time feedback.

Product Concept & Architecture

The Realtek RTL8763EW dual-core Cortex-M33 SoC integrates BLE 5.0, GPS baseband, and display driver. A Sony CXD5605 GNSS provides multi-constellation positioning. HRS3300 PPG measures heart rate. A 1.43-inch AMOLED display provides always-on visibility.

Engineering Solution

We developed a cascaded activity classifier: a lightweight decision tree (2KB) performs initial high-level classification (running vs. stationary vs. cycling) using accelerometer variance and GPS speed. A second-stage TinyML CNN processes 10-second windows of HR+IMU data to distinguish sub-types. Recovery score is computed from HRV decay rate post-exercise.

Result

93% automatic exercise classification accuracy across 20 sports. 14-day battery life with daily 1-hour GPS workout.