What happens once training stops: frozen weights, prefill vs decode, sampling, the KV cache, and the serving tricks that make inference faster without changing what the model knows.
How SFT and preference tuning (RLHF/DPO) turn a pre-trained next-token model into a helpful assistant: same training loop, different data, thin layer that shapes behavior not knowledge.
What happens inside the weight tensors as an LLM trains: random initialisation, the four-step training loop (forward, loss, backprop, optimizer), and what each component learns as a next-token model.
From public SaaS to air-gapped on-prem: a practical ladder of LLM hosting options, and how to choose based on regulatory pressure and engineering maturity.
Hyperparameters, vocabulary, and why a 7B model is 7B before day one, how the dials connect, where parameters live, and why the context window is a budgeted choice.
An intent router is the first deterministic decision in an agent stack. It maps user requests to the right workflow, agent, and tool manifest before the model loop runs — and when it fails, every downstream stage can execute perfectly and still miss the user’s goal.
Proposal is not permission. Agents propose tool calls; governance decides whether they run. An architecture breakdown of runtime trust boundaries for production agent systems in regulated industries.
Hallucination is not the model failing. It is the model succeeding at the wrong objective in a system that never gave it the right one.
A 20,000-ft view of the LLM lifecycle and why understanding the four stages matters for enterprise architecture.