I turn research prototypes into dependable engineering workflows for real technical data.
Industrial AI / Agent Systems / Case-Based Reasoning
Problem: Standard document retrieval surfaces passages without preserving the reasoning process required for technical tasks.
Contribution: Co-engineered procedure retrieval, objective evaluation, and fallback logic around the LLM, reducing unnecessary LLM usage through selective context and caching.
Industrial AI / Time-Series Machine Learning / Model Validation
Problem: CAE simulation assessment requires evaluating complex multichannel time-series signals alongside tabular test parameters.
Contribution: Learned expert assessment patterns from multichannel simulation signals and tabular context using modular time-series models and bootstrapped confidence intervals.
Industrial AI / Graph Neural Networks / Scientific ML
Problem: High-dimensional crash simulation requires rapid approximate predictions for mesh deformation and time-dependent acceleration.
Contribution: Developed and evaluated GNN surrogate workflows for selected crash-simulation outputs using diagnostic evaluation tools.
Industrial AI / MLOps
Problem: Crash-test sensor channel validation requires a deployable inference path combining signal imagery and tabular features.
Contribution: Built containerized AWS Lambda signal processing and classification workflows for crash-test sensor time-series data with Pydantic event validation.
An end-to-end computer vision pipeline for collecting, filtering, labeling, and classifying EV charger imagery.
A modular ML workflow for classifying time-series measurement channels as OK or NOK and supporting the review of abnormal signal patterns.
A conversational business-intelligence platform built around a stateful LangGraph ReAct assistant and deterministic retail analytics.
A lightweight Streamlit interface for following multi-agent financial analysis and reviewing generated reports.
A native macOS Skill and CLI for private speech recognition, translation, and bilingual subtitle generation.
A native macOS vision plugin that gives text-only DeepSeek Harness models local OCR and visual evidence.
A read-only Trading 212 portfolio workspace for dsh with live performance tracking, ECharts price history with buy/sell fills, and conversational tools.
Find roles in Germany, review the evidence for fit, and track applications in one workbench.
A macOS menu bar utility for monitoring DeepSeek API balance with secure local configuration.
Inverse and neural identification of anisotropic sheet-material parameters from a biaxial specimen.
Uncertainty-aware iterative tool correction for deep-drawn sheet-metal components.
A blood-cell analysis platform spanning segmentation, labeling, active learning, correction, and reporting.
An interactive single-image 3D reconstruction using vanishing-point geometry and a five-plane room model.
Retrieval, structured outputs, objective evaluation, observability, fallback behavior, and cost-aware orchestration.
Time-series signals, simulation data, anomaly detection, domain-aware evaluation, and uncertainty reporting.
CAE surrogates, geometry processing, diagnostic interfaces, reproducible experiments, and model integration.
My background in mechatronics and computational engineering led me from simulation and material modeling to building production AI systems.
Today I work across LLM workflows, industrial time-series data, simulation, and model evaluation, with a focus on systems that are reproducible, inspectable, and useful.