Parts
14 reusable parts cover all 666 use cases. Set a part up once, and every use case that needs it can share it, including the hardware it runs on.
LanguagePerceptionNumbers & decisionsAction & generationSpecialist
Language modelLLM & NLP inference, guardrails
Needed by 544 use cases. Smallest hardware: No GPU needed.
- Serving
- vLLM, SGLang, llama.cpp / Ollama (small setups)
- Open models
- Mistral Small 4 / Medium 3.5, Qwen3, gpt-oss-20b/120b, Llama, Gemma
- EU cloud
- Mistral La Plateforme (EU), Azure OpenAI EU Data Zone, AWS Bedrock Frankfurt, STACKIT AI Model Serving, IONOS AI Model Hub
- Gateway
- LiteLLM, Portkey
- Guardrails
- Llama Guard, NeMo Guardrails, Presidio (PII)
Agents & automationAgents, workflow automation & integration
Needed by 442 use cases. Smallest hardware: No GPU needed.
- Orchestration
- n8n (self-host), Camunda, Temporal
- Agent frameworks
- LangGraph, OpenAI Agents SDK, Claude Agent SDK, CrewAI, PydanticAI
- Protocols
- MCP, A2A
- Browser
- Playwright, Browser Use
Prediction & time seriesPredictive analytics: tabular, time series, anomalies
Needed by 312 use cases. Smallest hardware: No GPU needed.
- ML
- XGBoost / LightGBM / CatBoost, TabPFN, scikit-learn
- Forecasting
- Nixtla StatsForecast / TimeGPT, Chronos, TimesFM
- Anomaly
- PyOD, Isolation Forest, Anomalib
- Streaming
- Kafka, Flink
- MLOps
- MLflow, Feast
Document readingDocument parsing & vision-language
Needed by 283 use cases. Smallest hardware: 1 GPU workstation.
- Parsing
- Docling, PaddleOCR / PaddleOCR-VL, Unstructured, Marker
- Vision-language
- Qwen-VL, Mistral OCR, Gemini / GPT via EU region
Search & RAGRetrieval / RAG & enterprise search
Needed by 269 use cases. Smallest hardware: No GPU needed.
- Embeddings
- bge-m3, Qwen3-Embedding, multilingual-e5
- Stores
- pgvector, Qdrant, OpenSearch, Weaviate
- Rerankers
- bge-reranker-v2, Qwen3-Reranker
- Frameworks
- LlamaIndex, Haystack, LangChain
- Graph
- Neo4j, GraphRAG
VisionComputer vision & edge inference
Needed by 196 use cases. Smallest hardware: No GPU needed.
- Models
- YOLO (Ultralytics), RT-DETR, SAM 2, DINOv2
- Runtimes
- ONNX Runtime, OpenVINO, TensorRT / DeepStream, Hailo, Jetson
- Labelling
- CVAT, Label Studio
OptimizationOptimization, simulation & RL
Needed by 124 use cases. Smallest hardware: No GPU needed.
- Solvers
- OR-Tools, HiGHS, Gurobi, Timefold
- Simulation
- AnyLogic, SimPy, NVIDIA Isaac / Omniverse
- RL
- Stable-Baselines3, RLlib
SpeechSpeech & audio
Needed by 105 use cases. Smallest hardware: No GPU needed.
- Speech-to-text
- Whisper / faster-whisper, NVIDIA Parakeet/Canary, Voxtral
- Text-to-speech
- Piper, Kokoro, XTTS, ElevenLabs / Azure Speech (EU)
- Diarization
- pyannote
- Realtime
- LiveKit Agents, Pipecat
Coding assistantCoding assistants
Needed by 66 use cases. Smallest hardware: GPU server.
- Tools
- Claude Code, GitHub Copilot, Cursor, Continue.dev / Tabby (self-host)
- Open models
- Qwen3-Coder, Devstral, gpt-oss
Science & Earth dataScientific & Earth-observation foundation models
Needed by 46 use cases. Smallest hardware: GPU server.
- Models
- AlphaFold 3 / Boltz, ESM, Prithvi, Aurora / GraphCast
- Data
- Copernicus Data Space Ecosystem
RecommenderRecommendation & personalisation
Needed by 24 use cases. Smallest hardware: No GPU needed.
- Libs
- LightFM, Merlin, RecBole, two-tower retrieval on vector-db
RoboticsRobotics control
Needed by 22 use cases. Smallest hardware: 1 GPU workstation.
- Stack
- ROS 2, MoveIt 2, NVIDIA Isaac, LeRobot, VLA models (OpenVLA, π0)
BiometricsFace & biometrics (restricted)
Needed by 12 use cases. Smallest hardware: No GPU needed.
- Note
- Special-category data (GDPR Art. 9); some uses prohibited under AI Act Art. 5. Buy only from certified vendors; architecture-level guidance only.