AI

AI — currently my most active learning area. The notes here are mostly about LLMs, agents, and the glue that turns static models into useful systems. For a personal account of how I got here, see My AI journey.

Core concepts

• LLM — large language models, the substrate.
• AI agent — model + tools + control loop.
• Embedding — numeric representation of meaning.
• Reasoning — the thinking-token phenomenon in recent models.

Agent architectures and decision loops

• Multi-agent — multiple agents collaborating.
• Plan and act — separating planning from execution.
• Sequential thinking — forcing step-by-step reasoning.
• OODA, SOAR, system thinking — classical decision and cognitive frameworks that inform agent design.

RAG and retrieval

• RAG — retrieval-augmented generation, the pattern.
• Chunking — splitting documents before indexing.
• Unstructured-io — pre-processing messy inputs.
• Vector stores: Pinecone, Qdrant.

Frameworks

• LangChain, LangGraph, LangSmith — the LangChain family.
• LlamaIndex — focused on data/retrieval.

Agent protocols and tooling

• MCP — the protocol for exposing tools to agents.
• Cursor — my daily AI-native IDE. See the full Cursor MOC.
• Playwright MCP — browser automation as a tool.

Case studies

• Building an MCP app — draft on a recent MCP server/app project.
• One main agent per product — why I keep a single lead agent per project.
• Cursor MCP env vars — four ways to pass secrets to MCP servers.
• Playwright MCP with persistent session — keeping logged-in state across runs.
• OpenAPI → MCP → blog post — a content generation pipeline.