M.Sc. Mechatronics & Robotics (expected 12/2026) · Frankfurt University of Applied Sciences

I build embedded systems end to end, from the PCB to the firmware to the data coming out of it. Most of my work is C on nRF52 and STM32 (Zephyr, FreeRTOS), Python for ROS 2 and ML tooling, and TinyML with X-CUBE-AI.

Currently: Master's thesis at IoT Venture GmbH: firmware (nRF52840, Zephyr RTOS, IMU, BLE, SD logging) and a machine-learning pipeline for data-driven event detection in a connected bike tracker. Company project, so no code is published here.

Open to: full-time roles in Embedded Software / Firmware and Edge-AI from January 2027 (Rhine-Main area or remote).

Peripherals & protocols: DCMI · DMA · UART · I²C · SPI · CAN Bus · BLE
Tools: nRF Connect SDK / west · STM32CubeIDE · STM32CubeMX · X-CUBE-AI · MATLAB/Simulink · Autodesk Fusion / Eagle

Topics: Mobile robot control · Encoder odometry · PID control · Camera-based line following (team project) · TF2 · OccupancyGrid mapping

Topics: TinyML · MobileNet · IMU feature engineering · Random Forest (scikit-learn) · Post-training quantisation (PTQ) · On-device inference · ST Model Zoo

STM32H7 · FreeRTOS · OV2640 · X-CUBE-AI · PySide6

Object counting on an STM32H743ZI, compared two ways, both running on the chip: classical image processing (Otsu threshold, morphology, connected components) and a quantised TinyML model (MobileNetV1-0.25, 96×96 RGB, 279 KB Flash / 57 KB RAM). On 90 real-world test frames with changing light, TinyML reached 73.3 % accuracy and the classical pipeline 14.4 % (measured with a firmware version that has a known Otsu overflow bug, so this value is not a limit of the method and has not been re-measured); inference takes 59.8 ms vs. 66.6 ms. Course project (grade 1.0); AI assistance was used for the PySide6 PC dashboard, the ML training scripts and the code comments.

ROS 2 · Python · OpenCV · PID control · Dead-reckoning

Autonomous line-following and path-mapping for a Zumo robot. Camera-based line detection (team project; Kalman-filtered centroid in the debug overlay), PID motor control over serial, encoder-based dead-reckoning odometry, TF2 broadcasting, OccupancyGrid mapping, and a PyQt5 live dashboard (created with AI assistance).

C++ · Arduino Leonardo (ATmega32U4) · Interrupt-driven encoder · PD controller

Ported the Pololu Zumo 32U4 encoder library (MIT licence) to the Zumo Shield with an Arduino Leonardo (ATmega32U4). Interrupt-driven quadrature decoding with attachInterrupt() (channel A edges, direction from channel B), signed 32-bit atomic tick counters, and a discrete-time PD steering controller.

MATLAB/Simulink · CAN Bus · Vehicle Network Toolbox · DBC

Controlling a real BMW E9x instrument cluster over CAN with MATLAB/Simulink. Custom DBC file, signal generation in Simulink, and validation on the real cluster via the Vehicle Network Toolbox (over 90 % of the signals displayed correctly).

STM32F4 · PCB Design · DRV8848 · INA138 · MATLAB PID Tuner

Custom PCB for a real-time PID demonstrator: STM32F4-Discovery as controller, DRV8848 H-bridge, INA138 current sensing, USB-C Power Delivery input (12 V) with on-board step-down, and hardware-adjustable Kp/Ki/Kd via potentiometers. PCB v1 was built and tested; a revised v2 was designed but not manufactured. My part: hardware design and system integration; I supported the PID tuning with MATLAB PID Tuner.

📍 Frankfurt am Main, Germany