Hi, I'm Leon Schuster.

Compared Random Forest, Causal Forest, and a PyTorch neural network to predict machine failure on the AI4I 2020 predictive-maintenance dataset.

• Random Forest classifier reached 0.97 ROC-AUC on held-out data.
• SHAP and feature importance pinpointed torque, rotational speed, and tool wear as the dominant failure signals.
• Causal Forest estimated how machine quality tiers causally shift failure risk, beyond simple correlation.

Evaluated and optimized an image classification tool for recognizing automotive spare parts, comparing Vision Transformer and AWS Titan embeddings across preprocessing pipelines.

• Best setup — 1024-dim Titan embeddings, linear scoring, raw images — reached ~94% accuracy across 9 part categories.
• Added an "Unknown" category with distance-based filtering to flag out-of-domain images.
• Ran a cost analysis for a production AWS Bedrock / S3 Vectors deployment.

Investigated whether LLMs could automate extraction of technical vehicle specs from PDFs and web sources into TecAlliance's Global Editorial Tool for Vehicle Data, replacing a manual, error-prone data-entry workflow.

• Benchmarked 6 models — GPT-4o, o1, o3-mini, Gemini 2.0 Flash, DeepSeek R1, Copilot GPT-4 — on completeness, accuracy, and consistency.
• Found frequent inconsistent model codes/power figures and hallucinated values — manual validation is still required.
• Recommended Retrieval-Augmented Generation (RAG) and refined prompt engineering as the next step.

Implemented and evaluated anomaly detection methods to improve data quality on TecAlliance's B2B ordering platform, applying them to real order data to flag issues such as excessively large order quantities.

• Compared Z-Score, Isolation Forest, Boxplot, and SVM methods on real order data.
• Evaluated approaches on precision, computational cost, and scalability.