Ghost Edge AI Sticker

1. Introduction & Concept

The Ghost Edge AI Sticker is a paper-thin, flexible sticker-style smart sensor node built around the next-generation Nordic Semiconductor nRF54L15 wireless SoC.

Traditional IoT sensor nodes are often bulky, requiring rigid PCBs and battery cases that limit where they can be mounted. To break this limitation, we designed a device utilizing a Flexible PrintedCircuit (FPC) board.

This bendable, sticker-like design allows the node to conform to curved surfaces, making it ideal for unobtrusive physical activity tracking, structural health monitoring, and smart wearables.

Equipped with a 6-axis IMU (LSM6DSO) and a digital PDM microphone, this thin sticker runs real-time machine learning models (Edge AI) directly on-device using a dual-core processor architecture (ARM Cortex-M33 + RISC-V FLPR).

2. Key Features of this Project

This project main topics are as follows:

• FPC Design: Ultra-thin, flexible, and mountable on curved surfaces.
• Dual-Core Processing: ARM M33 and RISC-V core collaboration via IPC.
• Edge AI: Real-time IMU (and mic) data analysis using Edge Impulse.
• OTA Updates: Wireless firmware updates over BLE via MCUboot.

This bellow video summarizes the main results. It's all wrapped up in about two minutes, so please take a look!

3. Hardware Design

The custom FPC board layout and schematic were designed in KiCad. The sensor section is placed at the tip of a narrow flexible tail, allowing it to be positioned away from the main power and debugging circuitry.

• KiCad Design Files: Fully open-source project files for schematics and PCB/FPC layouts.

• Manufacturing Output: Gerber files and Bill of Materials (BOM) ready for FPC manufacturing.

• Schematic Drawings: PDF schematic sheets for quick hardware tracing.

https://github.com/iotengineer22/ghost-sticker/tree/main/board

4. Software & Firmware Architecture

All firmware applications are built using the nRF Connect SDK (NCS) / Zephyr RTOS and organized inside the /src folder:

https://github.com/iotengineer22/ghost-sticker/tree/main/src

src/
├── blinky/                  # Alternating dual LED test
├── dfu_test/                # BLE Peripheral UART with MCUboot and OTA DFU
├── peripheral_lbs/          # BLE LED Button Service (LBS) demo
├── peripheral_uart_i2c/     # LSM6DSO data BLE streaming (High-Speed Packed Mode)
├── peripheral_uart_i2c_demo/# LSM6DSO data BLE streaming (1Hz Demo Mode)
├── peripheral_uart_i2c_edge/# Edge Impulse gesture classification & BLE notify
├── peripheral_uart_pdm/     # Edge Impulse audio classification via PDM mic (WIP)
├── riscv_gpio/              # Multi-core IPC (ARM + RISC-V) LED controller
└── riscv_uart/              # Multi-core IPC (ARM + RISC-V) LSM6DSO monito

5. Multi-Core Inter-Processor Communication (IPC)

One of the highlights of this project is the multi-core utilization of the nRF54L15 SoC. Using the riscv_uart sample, the Remote core (RISC-V FLPR) handles high-frequency sensor acquisition from the LSM6DSO IMU over I2C.

The RISC-V core then transfers the formatted data packets to the Host core (ARM Cortex-M33) via IPC (using icmsg shared memory).

The Host core acts as the BLE gateway, receiving data over IPC and forwarding it to the connected BLE client using the Nordic UART Service (NUS). This separates the critical time-sensitive sensor.

6. Gesture Classification with Edge Impulse

By deploying the peripheral_uart_i2c_edge application, the nRF54L15 performs on-device real-time machine learning. We trained a gesture classification model (e.g., Wave, Shake, Idle) on Edge Impulse Studio using accelerometer data.

The C++ deployment library is integrated directly into the firmware. Classification outputs are printed in a clean single-line format:

Predictions (DSP: 4 ms, Class: 8 ms): idle: 0.02, wave: 0.98

This output is transmitted live over BLE, enabling low-latency control of smart devices using physical gestures.

7.Hardware Feedback (Lessons Learned)

During the hardware testing phase, we identified a footprint issue on the PCB/FPC artwork regarding the PDM microphone:

• Footprint Orientation Error: The PDM microphone component footprint was designed with a 180-degree rotation error (placed backward) on the board. Because of this footprint mistake, we were unable to debug or confirm the operation of the audio features on the actual physical hardware.
• Important Design Note: Although we have added warning notes to the uploaded schematic PDF and PCB layout artwork, anyone attempting to manufacture this board should carefully review the component datasheets and verify the footprint pins before starting production.

8. Conclusion & Links

This project main topics are as follows:

• FPC Design: Ultra-thin, flexible, and mountable on curved surfaces.
• Dual-Core Processing: ARM M33 and RISC-V core collaboration via IPC.
• Edge AI: Real-time IMU (and mic) data analysis using Edge Impulse.
• OTA Updates: Wireless firmware updates over BLE via MCUboot.

This bellow video summarizes the main results. It's all wrapped up in about two minutes, so please take a look!

This has been a great fun challenge.

Thanks to NextPCB for hosting this exciting program. Thank you very much for all the support provided.

https://jp.nextpcb.com/blog/build-with-riscv-freedom