Alle Pakete

The core AI safety toolkit: alignment and RLHF, bias and fairness auditing, red-teaming and adversarial testing, robustness to distribution shift and adversarial inputs, and the governance frameworks that shape responsible deployment.

Claude API, prompting, tool use, agents, and the Anthropic ecosystem.

Build production apps on the Claude API: the Messages API request/response shape, prompt caching for cost and latency, streaming responses, tool use and function calling, and vision/multimodal inputs.

ML fundamentals, neural networks, LLM fine-tuning, and Andrew Ng's curriculum.

When and how to adapt a model instead of just prompting it: SFT and instruction tuning, LoRA/QLoRA/PEFT, dataset curation, training dynamics, preference optimization (RLHF and DPO), evaluation against the base model, quantization for serving, and the cases where fine-tuning does not help — model-agnostic.

Vertex AI, Gemini, cloud ML services, and Google's AI certification path.

Watson, watsonx, IBM AI fundamentals, and enterprise AI patterns.

Compose LLM applications with LangChain: LCEL pipes and the Runnable protocol, chains (RAG, structured output, fallbacks), tool-calling and ReAct agents, chat history and memory, and LangGraph nodes/edges for stateful, resumable agent graphs.

Turn a model into an agent: tool/function calling, the ReAct loop, planning, memory, multi-step execution, multi-agent patterns, guardrails, and evaluating agent reliability — vendor-neutral.

Call LLMs over HTTP well: chat messages and roles, streaming, tokens and pricing, context limits, temperature/top_p and stop sequences, structured output, tool calling, rate limits, retries and timeouts, idempotency, caching, batching, latency, and the embeddings endpoint.

Measure whether an LLM system actually works: eval sets, task metrics and their limits, LLM-as-judge, human evaluation, regression testing, faithfulness and hallucination checks, red-teaming, safety and bias, offline vs online, and reproducibility.

Operate LLM applications in production: deployment patterns, prompt versioning, monitoring and tracing, cost control, guardrails and PII handling, evals in CI/CD, canary releases and rollback, caching, scaling, fallbacks and model routing, security, rate limiting, and data-retention compliance.

Azure AI, Copilot, Microsoft Certified: AI Fundamentals (AI-900), and Azure OpenAI Service.

The free on-ramp to working with LLMs: instructions, roles, few-shot, chain-of-thought, structured output, context windows, grounding, and evaluating prompts — model-agnostic.

Ground an LLM in your own data: chunking, embeddings, vector search, top-k retrieval, reranking, hybrid search, context assembly, citations, and how to evaluate a RAG system.