The Words You Will Hear But Never UsePretraining, RLHF, LoRA, overfitting, distillation. This course skips all of it on purpose, and you still have to know what the words mean. Here is the vocabulary of the world next door, and how far each term sits from your actual job.evergreen · 29 July 2026Running Your Own ModelOpen weights make self-hosting possible. A GPU bill, an on-call rotation and a licence you have to actually read decide whether it is wise. Here is when running your own model is the right call, and when it is a hobby with a budget.evergreen · 29 July 2026Caching Is the Cheapest WinBefore you switch models or tune a serving engine, look at what you resend on every call. An agent repeats the same prefix hundreds of times, and caching it is the one cost lever that costs nothing in quality.evergreen · 29 July 2026What You Must Not SendGuardrails protect your agent from the outside. This is the other direction: what leaves your machine on every call, who stores it, in which country, and the data you must never put in a prompt.evergreen · 29 July 2026What MCP Actually StandardisesMCP is described as USB for tools, and that is true of the plug. Underneath it standardises three things a model can be given, two ways to connect, and one handshake. Everything it does not standardise becomes your problem.evergreen · 29 July 2026Reasoning and Test-Time ComputeA reasoning model buys accuracy with time. It generates thinking tokens before answering, which you pay for and rarely see. Knowing when that trade is worth it, and when it is waste, is now a routine engineering decision.evergreen · 29 July 2026What Models Take InA model does not see your image. It receives patches turned into tokens, in the same stream as your text. Once you know that, image cost, resolution limits and the choice between OCR and native vision stop being mysteries.evergreen · 29 July 2026Shipping Agents Is a Level of Its OwnMost agent projects die between the demo and the deploy. Shipping agents to production is its own engineering discipline: serving, cost, observability, and durable execution. Here is the last rung of the ladder.evergreen · 11 July 2026An Agent You Cannot Measure Is an Agent You Cannot ImproveMeasuring agents is the skill that separates a demo from a product. It looked fine when I tried it is not measurement, and this is how you replace that feeling with numbers you can trust.evergreen · 11 July 2026Agents You Can Trust Are Built From GuardrailsA demo that works once is not a product. Agent security is the discipline of bounding what an agent can do: guardrails, prompt injection defence, least privilege, and graceful failure.evergreen · 11 July 2026When One Agent Is Not EnoughMost problems never need a multi-agent system. Here is the honest cost of a swarm, when you actually earn one, and the coordination tax nobody prices in.evergreen · 11 July 2026Building a Single Agent Is a Loop With a GoalA single agent is not clever wiring. It is one loop that plans, acts, remembers, and corrects toward a goal. Its quality is decided by one thing: what you let into the context window.evergreen · 11 July 2026Tools Turn a Chatbot Into an ActorA chatbot answers. An agent acts. The one thing between them is a tool. This is how giving models tools works: schemas, the agentic loop, routing, and the MCP standard that kills bespoke glue.evergreen · 11 July 2026RAG Is a Search Problem Wearing an AI CostumeRAG lets a model answer from your data without retraining. But RAG is a search problem in disguise, and it fails exactly where search fails, not where AI fails. Here is the whole retrieval pipeline, including the parts everyone skips.evergreen · 11 July 2026Prompt Engineering Is Software EngineeringPrompt engineering looks like writing to a chatbot. The moment your code has to parse the reply, it becomes software engineering: system prompts, JSON Schema, function calling, and retry loops.evergreen · 11 July 2026A Language Model Is a Probability MachineHow a language model works comes down to one fact: it predicts the next token. Tokens, embeddings, attention, context, and every sampling knob you touch all live inside that single trick.evergreen · 11 July 2026The Applied AI Engineering LadderAI engineering is not machine-learning math. It is a ladder of nine engineering skills, and every level assumes the one below it. Skipping rungs is why most agent projects fail.growing · 11 July 2026What an LLM Actually IsA language model does one thing. It predicts the next token. It is frozen, it has no memory, and it cannot act. Everything else you will build in this course is a ring around that one function.evergreen · 30 June 2026 All writing