At Ruhr University Bochum, I implemented a deep Q-network with a CNN encoder from scratch in NumPy, including its forward and backward passes and Adam optimiser, then verified it with numerical gradient checks.
I built reproducible Python workflows for simulation, training, and automated evaluation, with controlled multi-seed experiments, logged configurations, checkpoints, data-quality checks, and error analysis. I also developed and trained a closed-loop spiking neural network agent using NEST, reward-modulated learning, and memory replay.
At Max Planck Institute MiS, I ran large-scale simulations and statistical analyses of noisy, high-volume data, contributing to three peer-reviewed papers. In engineering projects, I packaged an ML pipeline with leakage-free evaluation and tests, and built an LLM agent service with a self-built MCP server and code-enforced safeguards.

