Expert Track

Attention, the KV cache, positional encoding and mixture-of-experts — the architectural facts that explain why inference costs and behaves the way it does.

What changed when models started thinking before answering, how to decide when the extra tokens pay for themselves, and why visible reasoning is not a trace.

System prompt architecture, constrained decoding, injection defence and the discipline of deciding what goes into the context window.

Where the money actually goes in an LLM system, how prefix caching works and silently breaks, and the order to pull cost levers in.

Chunking, hybrid retrieval, reranking and the metrics that tell you which half of your pipeline is broken.

The agent loop as a system to be bounded, tool design that survives contact with a model, and the multi-agent patterns that actually earn their complexity.

Building an evaluation harness that catches regressions, the documented biases of model graders, and how to run evals in CI without them becoming theatre.

When fine-tuning is the right instrument, how LoRA and quantisation make it affordable, and why dataset quality decides the outcome.

Reliability, observability, failure isolation and rollout discipline — the engineering that turns a working prototype into a system you can operate.