# Marius Schlüter — Engineering Portfolio Curiosity drives me to explore new technologies and put them to work. I build across AI, robotics, optimization, and industrial software—from LLM-controlled robots and a 200,000-parameter solar-car optimizer to .NET and Angular applications deployed on machines worldwide. Location: Munich, Germany ## Contact - Email: [themrslue@googlemail.com](mailto:themrslue@googlemail.com) - Website: [https://kmode.dev](https://kmode.dev) - GitHub: [https://github.com/kmodexc](https://github.com/kmodexc) - LinkedIn: [https://linkedin.com/in/mariusschlueter](https://linkedin.com/in/mariusschlueter) ## Solar-car driving strategy Category: Large-scale optimization Context: Head of Driving Strategy, TUfast e.V. · May 2024–present A solver-agnostic system that calculates how a solar car should use limited energy during the Bridgestone World Solar Challenge. - **Optimization:** Developed an AI-based race simulation with one-second accuracy and solved a constrained nonlinear problem with 200,000 parameters. - **Compute:** Used PyTorch, Adam-style optimization, CUDA Graphs, TensorBoard, Docker, CI/CD, and NVIDIA H100 GPUs. - **Operations:** Led strategy simulation and supported remote race telemetry with PostgreSQL, Grafana, VPN/Tailscale, and Starlink. Technologies: PyTorch, CUDA, Adam2, Docker, PostgreSQL, Grafana Project images: -  — The solar car whose race energy use the strategy system optimizes. ([Full size](assets/images/lux025.png)) -  — Technical output across velocity, battery, solar power, and race constraints. ([Full size](assets/images/sao-output-graphs.png)) ## Confidence calibration of 3D LiDAR object detectors Category: Machine-learning research Context: Master’s thesis, TUM / MIRMI · June 2024–January 2025 · Grade 1.7 Research into whether a 3D detector’s confidence scores reliably represent the probability that its predictions are correct. - **Result:** Reduced miscalibration by 12.8 percentage points, from 15.9% to 3.1%, using custom loss functions and calibration tooling. - **Research:** Trained and evaluated 3D LiDAR detectors on the Waymo dataset using NVIDIA H100 cluster resources. - **Publication:** The resulting paper was accepted at the 2025 IEEE Intelligent Vehicles Symposium. Technologies: Python, PyTorch, Waymo dataset, Slurm, Docker, NVIDIA H100 Links: - [Read the IEEE paper](https://doi.org/10.1109/IV64158.2025.11097526) ## TUfast autonomous stack Category: Autonomous systems Context: Autonomous development, TUfast e.V. · May 2022–May 2024 Perception and sensor integration for a Level 4 competition vehicle in the Shell Eco-marathon Autonomous Urban Concept category. - **Perception:** Built 3D localization, map-less driving, OctoMap mapping, camera/LiDAR fusion, and road and parking-spot detection. - **Integration:** Developed a custom Sekonix camera driver, merged camera and LiDAR point clouds, used Coral TPU acceleration, and prepared NVIDIA DRIVE. - **Leadership:** Led the autonomous team from May 2023; trained detectors on a self-developed dataset using LRZ AI Cluster resources. Technologies: ROS2, Python, C++, PyTorch, TensorFlow, OctoMap, Coral TPU Links: - [Watch the vehicle video](https://www.youtube.com/watch?v=tYWKW2fQkIY) Project images: -  — The system layout connects vehicle hardware, sensors, and compute components. ([Full size](assets/images/pic-autonom.png)) ## TUfast matching tool Category: Combinatorial optimization Context: TUfast engineering software · Public repository A desktop planning tool that turns constrained shift assignment into a graph-optimization problem. - **Problem:** Assign shift leads and workers while respecting planning constraints. - **Approach:** Modeled constraints with maximum-flow and min-cost-flow algorithms using NetworkX and SciPy sparse graphs. - **Delivery:** Implemented and tested the matching behavior, then packaged it as a PySide6 GUI application. Technologies: Python, NetworkX, SciPy, PySide6, Max flow, Min-cost flow Links: - [View source on GitHub](https://github.com/kmodexc/matching_tool/tree/main) Project images: -  — The delivered desktop application collects planning inputs and generates a constrained shift plan. ([Full size](assets/images/matching_tool.png)) ## End-to-end LTE/4G vehicle telemetry Category: Embedded infrastructure Context: TUM School of Computation, Information and Technology · October 2023–May 2024 Reliable, energy-efficient telemetry connecting the muc023+ competition vehicle to remote engineering tools. - **Edge:** Collected vehicle data with Linux, Raspberry Pi, CAN bus, C/C++, and Python. - **Transport:** Connected the vehicle over LTE/4G using VPN and Tailscale. - **Backend:** Used PostgreSQL database replication and a Grafana web interface for remote monitoring. Technologies: Linux, Raspberry Pi, CAN bus, LTE/4G, C/C++, PostgreSQL, Grafana ## Self-learning dough-strand controller Category: Industrial AI Context: Bachelor’s thesis with Lieken GmbH · June–August 2021 A self-learning controller that keeps bread weight constant as dough consistency changes on an automated production line. - **Control:** Designed and optimized the controller around two key variables in the automated dough-flow process. - **Delivery:** Implemented the self-learning controller in Python for an industrial production setting. Technologies: Python, PLC, ODBC, pytest, PyPI, GitHub Actions ## Provenance Selected from `../project_portfolio.tex`; factual authority: `../CV_INFORMATION.md`. The website and this document are generated from `content.json`.