Shipping AI that solves real work: agents, tools/MCP, structured outputs, human-in-the-loop, observability, and choosing the smallest reliable stack over hype.

Working with private datasets safely: data hygiene, PII, RAG vs fine-tuning vs adapters (LoRA), eval sets, and keeping proprietary knowledge out of public model training.

Running models locally for privacy and cost control (Ollama, LM Studio, llama.cpp). Hardware trade-offs, quantization, offline RAG, and when local beats cloud APIs.

Short- and long-term agent memory: conversation state, summary memory, entity/user profiles, tool result caching, and when to persist vs. re-retrieve from a knowledge base.

Retrieval-Augmented Generation: chunking, embeddings, vector stores, hybrid search, reranking, citation, and grounding LLMs on private docs without leaking source data.

Hands-on with GPT APIs, Assistants/tools, embeddings, and production patterns: rate limits, cost control, safety, and integrating models into WordPress/Laravel apps.

Practical use of Claude for coding agents, specs, refactoring, and tool-using workflows. Prompt design, evals, and shipping AI-assisted features in real products.