Frameworks: PyTorch (Standard), Hugging Face Transformers, scikit-learn, XGBoost, LightGBM, CatBoost, Keras / TensorFlow bei Bedarf
Fine-Tuning: Hugging Face TRL, Axolotl, Unsloth, PEFT (LoRA, QLoRA), DeepSpeed, FSDP, vLLM für das Serving
Computer Vision: YOLOv8/v10, Ultralytics, MMDetection, Detectron2, SAM 2, OpenMMLab, OpenCV, Roboflow für Dataset-Ops
Experiment-Tracking: MLflow, Weights & Biases, Neptune, Comet, ClearML – versionierte Datensätze, Modelle und Runs
Daten und Feature Stores: Pandas, Polars, DuckDB, Dask, Spark; Feast, Tecton für Feature Stores; DVC für Dataset-Versionierung
Trainingsinfrastruktur: AWS SageMaker, Vertex AI, Azure ML, Modal, Lambda Labs, RunPod, Paperspace, On-Prem A100/H100
Inference Serving: BentoML, Triton Inference Server, TorchServe, vLLM, TGI, Modal, AWS Lambda + Container, SageMaker-Endpunkte
Edge / On-Device: Core ML, TFLite, ONNX Runtime, MLC LLM, Llama.cpp für mobile und eingebettete Inferenz
MLOps & Monitoring: Evidently AI, Arize, WhyLabs, Fiddler – für Drift-, Datenqualitäts- und Bias-Monitoring in der Produktion
Evaluation & verantwortungsvolle KI: Individuelle Evaluations-Harnesses, Model Cards (Hugging Face / Google), Fairness-Metriken aufgeschlüsselt nach demografischen Segmenten