Resource shelf

The inference runtime this lab is built around. Runs quantised models on modest hardware and exposes the knobs that actually matter.

Where the weights live. Read the model card and the licence before you read the benchmark numbers.

The current baseline family here. Good size coverage, and the small variants are unusually capable for what they cost in memory.

The fastest path from nothing to a model answering questions. Worth starting here before dropping down to raw llama.cpp.

A desktop front end for local models. Useful for quick comparisons when you do not want to script anything.

Where you go when serving throughput starts to matter more than fitting on one consumer card.

An open protocol for connecting models to tools and data. Worth understanding structurally, not just as an integration.

Microsoft's orchestration SDK for agents and plugins. Sits naturally alongside a Graph-based tool surface.

Graph-shaped agent workflows with explicit state. Useful when the control flow is the thing you need to reason about.

Tool use, context handling, and agent patterns, with unusually candid guidance on failure modes.

The Top 10 for LLM applications, plus the agentic security work. The most useful shared vocabulary available right now.

ATT&CK's structure applied to attacks on machine-learning systems. Good for turning a vague worry into a testable technique.

The governance layer. Dry, and the right thing to point at when someone asks how risk is being managed rather than mitigated.

Microsoft's automation framework for probing generative AI systems. Turns ad-hoc red teaming into something repeatable.

A vulnerability scanner for language models. Broad probe library, easy to point at a local endpoint.

Still the reference every detection gets mapped back to. Read the data sources section, not just the technique list.

Small, scoped technique tests. The generator half of any honest detection bench.

Vendor-neutral detection rules. Useful as a source of logic even when you write the final query in KQL by hand.

Analytics rules, hunting queries, and workbooks in the open. A good place to see how a rule is structured before writing your own.

The full query language, not the cheat sheet. Worth an afternoon — most slow queries are a missing operator away from being fast.

Identity governance, conditional access, and lifecycle workflows. Dense, and the primary source for anything identity-shaped.

How most automation in this lab reaches a tenant. Check the permission scope each cmdlet actually requires — the names are optimistic.

Deployment guidance rather than slogans. Most useful as a checklist against an environment you already have.

A vendor-neutral maturity view. Handy when you need to describe where an organisation actually sits.