Yunus Emre Dönmez

Efficient speech inference with lookup table networks (SNN-LUT). • Implemented the lookup table (LUT) layer from Izhikevich's 2025 Spiking Manifesto in PyTorch with custom autograd, turning matrix multiplications into comparisons and table reads. • Distilled a pretrained SpeechBrain keyword spotting model (96.4% on Google Speech Commands v2) into 512 LUTs, cutting model level memory traffic by 91% while keeping 90% of the accuracy. • Served both models from NXP's NAVQ+ board through an HTTP inference API w

Quadrotor Navigation using Neuromorphic Control Policies (graded 1.3) • Built a drone navigation framework in PyBullet simulation where a small neural policy corrects a classical controller while the quadrotor flies through gates in random wind. • Trained 132 parameter spiking neural networks with evolution strategies (CMA-ES) and benchmarked them against ANN and reinforcement learning baselines (PPO, VPG) on 1,000 held out runs. • Compared 3 neuron models and 4 input encodings, tracking spike rate and syn

Working student role on embedded machine learning for capacitive touch detection (TouchSense) and IMU motion learning. • Owned the touch detection ML pipeline for capacitive sensor data, from data collection and Center of Mass label refinement to training, evaluation and deployment. • Shrank the model from 117k to 30k parameters (74% smaller) through hyperparameter optimization, while improving detection accuracy by 8%. • Quantized FP32 models to INT8 TensorFlow Lite, fixed a batch dimension issue in the I

Working student role. Built a gamified data collection system that improved the training data for ML based touch interaction models. • Created a touchscreen game in Python with a live timer and leaderboard that turned repetitive touch recordings into a challenge, with an RGB depth camera capturing ground truth. • Set it up at several places around the office, including the canteen, so colleagues played during breaks. This made data collection possible at scale. • A/B tests against the old setup showed 43%

• Built feature engineering and backtesting for gradient boosting models (XGBoost, scikit-learn) on financial time series. • Processed global market datasets and validated trading signals with statistical tests before backtesting. • Joined after a technical evaluation on gradient boosting models for financial time series prediction. Python · XGBoost · scikit-learn · Time Series Analysis · Backtesting

• Contributed to the development of a next-generation tachometer system for high precision telemetry measurement. • Designed and implemented a PyQt based graphical user interface (GUI) to support hardware configuration, testing workflows, and system validation. • Developed Python based parsing pipelines to decode raw hexadecimal telemetry streams, transforming low level binary data into structured, protocol compliant formats for downstream processing. • Engineered high value telemetry features to improve d

• Contributed to performance evaluation of next-generation photovoltaic (PV) systems, supporting structured testing and comparative efficiency assessments. • Assisted in analyzing operational data to evaluate panel output, stability, and environmental response characteristics. • Conducted field exposure at biogas production facilities, studying anaerobic digestion processes, biomass conversion mechanisms, and energy yield optimization principles.

Master's thesis at NXP Semiconductors on neuromorphic drone control: Quadrotor Navigation using Neuromorphic Control Policies, graded 1.3. Focus: Machine Learning, Deep Learning, Computer Vision. Courses include Fundamentals of Foundation Models and Deep Learning and Inverse Problems (Prof. Reinh