The Hackster.io project page describes the Ghost Edge AI Sticker as a paper-thin, flexible sticker-style smart sensor node built around the Nordic Semiconductor nRF54L15 wireless SoC. The board uses a Flexible Printed Circuit (FPC) design, integrates a 6-axis IMU (LSM6DSO) and a digital PDM microphone, and runs on a dual-core architecture pairing an ARM Cortex-M33 host with a RISC-V FLPR remote core. Firmware includes over-the-air updates via BLE using MCUboot, real-time gesture classification via Edge Impulse, and open-source KiCad files, Gerbers, and a BOM. The author flags a 180-degree PDM microphone footprint rotation error that prevented audio feature validation on physical hardware - a practical caveat for anyone planning to reproduce the design.
What happened
The Hackster.io project page published June 20, 2026, documents the "Ghost Edge AI Sticker" by hardware engineer misoji engineer (iotengineer22): a paper-thin, flexible sticker-style smart sensor node built around the Nordic Semiconductor nRF54L15 wireless SoC. The project is fully open-source with KiCad schematics, Gerber manufacturing output, a Bill of Materials, and firmware in a public GitHub repository.
Hardware design
The board is laid out as a Flexible Printed Circuit (FPC), with the sensor section placed at the tip of a narrow flexible tail to physically separate sensing from power and debug circuitry. The nRF54L15 dual-core architecture is used deliberately: the RISC-V FLPR remote core handles high-frequency IMU acquisition from the LSM6DSO 6-axis accelerometer+gyroscope over I2C, while the ARM Cortex-M33 host core acts as the BLE gateway, receiving formatted data packets via shared-memory IPC (icmsg) and forwarding them over the Nordic UART Service (NUS). Firmware is built on nRF Connect SDK and Zephyr RTOS, with BLE OTA updates via MCUboot. The nRF54L15 provides 2.5 MB flash and 1 MB RAM on a 22 nm process.
Edge AI
The Edge Impulse integration deploys a gesture classification model (Wave, Shake, Idle) trained on accelerometer data from Edge Impulse Studio. The compiled C++ inference library integrates directly into firmware. Output is a single-line prediction including DSP and classification latency: Predictions (DSP: 4 ms, Class: 8 ms): idle: 0.02, wave: 0.98. Results stream live over BLE for low-latency gesture-controlled device interactions.
Hardware lesson
During physical testing, a 180-degree footprint rotation error on the PDM microphone prevented audio feature validation on the manufactured board. The author added warning notes to the schematic PDF and PCB layout and advises manufacturers to verify component datasheets against footprint libraries before production. Audio classification firmware (peripheral_uart_pdm) is included in the repository but marked WIP due to this unresolved defect.
Context
The project was developed under the NextPCB hardware accelerator program, which provided FPC fabrication support. Edge Impulse added official support for the nRF54L15 development kit in 2025, covering motion recognition, keyword spotting, and audio classification workflows.
What to watch
Teams evaluating the nRF54L15 for wearable or conformable edge-AI products will find the IPC split - RISC-V FLPR for sensor acquisition, ARM M33 for BLE gateway - to be a reproducible, low-overhead pattern. The open KiCad FPC layout lowers the barrier for small-batch flexible PCB prototyping, though the PDM footprint incident underscores the importance of datasheet verification before tape-out.
Scoring Rationale #
A concrete open-source reference design combining FPC form factor, nRF54L15 dual-core IPC, and Edge Impulse inference provides solid practitioner value for teams prototyping wearable or conformable edge-AI nodes. The project breadth - KiCad FPC layouts, multiple Zephyr firmware examples, BLE OTA, and gesture classification - makes it useful across the embedded AI stack, though the PDM footprint defect and single-developer hobbyist scope limit this to a niche reference rather than a broadly impactful release.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.