Mechatronics engineer at Bosch, Stuttgart. I spent ~2 years managing requirements on ADAS projects, and since 2025 I build ML models. I still test them the way I tested requirements: every claim needs evidence.
A score tells you how good a model is. Its failures tell you whether you can ship it.
- 🔬 Now: in Bosch's AI research department, working on multi-agent systems.
- 🎯 What I bring: requirements discipline for ML: traceable results, tested pipelines, documented failure modes.
- 🤖 Side projects: LLM agents with verification built in: LangGraph multi-agent graphs, MCP tools, reviewer loops.
- 🔍 How I work: each project answers a question and shows where the model breaks, not just its best score.
- 🦾 Roots: robot arms, mobile robots, embedded motor control. I like software that has to work in the physical world.
- 🧭 Open to: AI and ML engineering roles in robotics, aerospace and defense.
 |
Latest Can an LLM agent correctly report where an object detector fails? Three specialist agents analyse YOLOv8n's failures; a reviewer checks every claim against the data and sends wrong sections back. LangGraph MCP YOLOv8 Ollama Docker
|
driving-scene-segmentation: how small can a segmentation model get before it is unsafe?
DeepLabV3+ 72.9% vs SegFormer-B0 60.1% mIoU on Cityscapes; trucks get labelled as cars. PyTorch Hugging Face
rul-for-nasa-jet-engines: how many cycles until a jet engine needs maintenance?
Random Forest on NASA C-MAPSS, RMSE 46 cycles, 4x better than linear baselines. scikit-learn pandas
medvision: how far does a small CNN get on colorectal histology?
EfficientNet-B0, 99.6% on a held-out split of NCT-CRC-HE-100K. PyTorch timm
ML and vision
LLMs and agents
MLOps and engineering
