Hi there, I am Niklas.
I like figuring out how things work.

A comprehensive three-part blog series exploring the development of a deep learning-based autoencoder for speech enhancement. Achieved significant improvements over reference methods with 69% increase in PESQ (Perceptual Evaluation of Speech Quality) and 9% increase in STOI (Short-Time Objective Intelligibility). Covers architecture design, training strategies, and deployment considerations.

Shaped the technical foundation as founding engineer for a mobility startup from zero to seed. Developed an automated market-intelligence pipeline to monitor market trends, enabling robust cost forecasts and strategic insights. Followed the ride for 9 months from zero to seed.

A modularized DRL framework for autonomous vessel control on inland waterways, featuring separate agents for high-level path planning (considering traffic rules and dynamic obstacles) and low-level path following. Outperforms traditional APF and PID controllers by 65% in obstacle avoidance while reducing control effort.

An open-source Python framework for extracting clean ship trajectories from noisy AIS big data. Uses maneuverability-dependent α-quantile-based filtering to handle technical inaccuracies and compliance issues in raw AIS records. Robustly extracts long, uninterrupted trajectories for maritime domain awareness and algorithm testing.

A bootstrapped Deep Q-Network (DQN) controller for autonomous vessel navigation on restricted inland waterways. Handles challenging conditions like high flow velocities and shallow banks on the Rhine. Trained in diverse, realistic environments and validated against real-world data, demonstrating superior adaptability compared to vessel-specific PID controllers.