A living technical feed and digital laboratory — AI, software, cyber, homelab, risk, MMA and anime.https://www.andraewilliams.com/https://www.andraewilliams.com/posts/evaluation-how-do-you-know-it-worked/https://www.andraewilliams.com/posts/evaluation-how-do-you-know-it-worked/"Looks right" and "is right" are different claims — an agent only checked by eyeballing the transcript will drift for weeks before anyone notices.Wed, 19 Aug 2026 00:00:00 GMTEXPERIMENTaihttps://www.andraewilliams.com/posts/guardrails-stopping-the-agent-from-hurting-itself/https://www.andraewilliams.com/posts/guardrails-stopping-the-agent-from-hurting-itself/Capable and safe to leave unattended are different claims. Guardrails close the gap — permission boundaries, confirmation on the irreversible, and a cap on how much one bad turn can do.Wed, 19 Aug 2026 00:00:00 GMTARCHITECTUREaicyberhttps://www.andraewilliams.com/posts/homelab-storage-you-can-afford-to-lose/https://www.andraewilliams.com/posts/homelab-storage-you-can-afford-to-lose/The lab rule inverted: decide what is disposable first, protect the tiny remainder properly, and skip the NAS you were about to buy.Wed, 19 Aug 2026 00:00:00 GMTTUTORIALhomelabhttps://www.andraewilliams.com/posts/how-ai-agent-memory-actually-works/https://www.andraewilliams.com/posts/how-ai-agent-memory-actually-works/Memory in AI agents isn't one thing. It's a stack of storage, indexing, and retrieval systems that fail in different ways.Wed, 19 Aug 2026 00:00:00 GMTARCHITECTUREaihttps://www.andraewilliams.com/posts/how-anime-fakes-speed-without-moving-the-camera/https://www.andraewilliams.com/posts/how-anime-fakes-speed-without-moving-the-camera/A held frame, held long enough to be uncomfortable, reads as faster than a camera that actually pans — because your eye is doing the moving instead.Wed, 19 Aug 2026 00:00:00 GMTANIMEanimehttps://www.andraewilliams.com/posts/how-judges-actually-score-a-close-round/https://www.andraewilliams.com/posts/how-judges-actually-score-a-close-round/The 10-point-must system rewards a specific, narrow definition of dominance — and most disputed decisions come from fighters optimizing for a different one.Wed, 19 Aug 2026 00:00:00 GMTMMAmmahttps://www.andraewilliams.com/posts/mfa-fatigue-the-push-notification-weak-link/https://www.andraewilliams.com/posts/mfa-fatigue-the-push-notification-weak-link/The control that was supposed to stop credential stuffing has its own failure mode — and it's a human one, not a cryptographic one.Wed, 19 Aug 2026 00:00:00 GMTTLDRcyberhttps://www.andraewilliams.com/posts/observability-watching-an-agent-think/https://www.andraewilliams.com/posts/observability-watching-an-agent-think/Part four called episodic memory the layer most agents skip. This is that layer, built — every decision logged, because a silent failure is the only kind you can't debug.Wed, 19 Aug 2026 00:00:00 GMTBUILDaihttps://www.andraewilliams.com/posts/picking-homelab-hardware-without-wasting-money/https://www.andraewilliams.com/posts/picking-homelab-hardware-without-wasting-money/The used-enterprise trap, the mini-PC sweet spot, and why your first server should cost less than your monitor.Wed, 19 Aug 2026 00:00:00 GMTTUTORIALhomelabhttps://www.andraewilliams.com/posts/pihole-ad-blocking-for-the-whole-house/https://www.andraewilliams.com/posts/pihole-ad-blocking-for-the-whole-house/One DNS server, every device on the network, zero browser extensions — and why this is the workload that actually convinces skeptics.Wed, 19 Aug 2026 00:00:00 GMTTUTORIALhomelabhttps://www.andraewilliams.com/posts/planning-and-reasoning-when-one-call-isnt-enough/https://www.andraewilliams.com/posts/planning-and-reasoning-when-one-call-isnt-enough/One tool call at a time works until the task has more than one step with a dependency between them. Then the loop needs an explicit plan, not just a next guess.Wed, 19 Aug 2026 00:00:00 GMTARCHITECTUREaihttps://www.andraewilliams.com/posts/proxmox-in-one-evening/https://www.andraewilliams.com/posts/proxmox-in-one-evening/Install the hypervisor first, even for one machine — snapshots make every later experiment reversible, which is the whole point of a lab.Wed, 19 Aug 2026 00:00:00 GMTTUTORIALhomelabhttps://www.andraewilliams.com/posts/shipping-an-agent-that-survives-contact-with-reality/https://www.andraewilliams.com/posts/shipping-an-agent-that-survives-contact-with-reality/Nine parts from the bare loop to a system with real memory, evaluation and guardrails — what actually mattered, and the one layer that made every other layer debuggable.Wed, 19 Aug 2026 00:00:00 GMTBUILDaihttps://www.andraewilliams.com/posts/tailscale-reach-your-lab-from-anywhere/https://www.andraewilliams.com/posts/tailscale-reach-your-lab-from-anywhere/Flat, private networking to every lab machine with zero opened ports — and why port forwarding was never the answer.Wed, 19 Aug 2026 00:00:00 GMTTUTORIALhomelabcyberhttps://www.andraewilliams.com/posts/the-agent-loop/https://www.andraewilliams.com/posts/the-agent-loop/Call the model, parse a tool call, run it, feed the result back. The whole mechanism takes an afternoon — everything else in this series is what happens after.Wed, 19 Aug 2026 00:00:00 GMTELI5aihttps://www.andraewilliams.com/posts/the-difference-between-a-risk-and-a-control/https://www.andraewilliams.com/posts/the-difference-between-a-risk-and-a-control/A risk register full of controls with no named risk behind them is the most common failure mode in risk management, and it's invisible until an incident asks the question directly.Wed, 19 Aug 2026 00:00:00 GMTELI5riskhttps://www.andraewilliams.com/posts/the-lab-eight-parts-later/https://www.andraewilliams.com/posts/the-lab-eight-parts-later/What actually got built, what it cost, and the two incidents that proved the 'allowed to break' rule from part one was the right one.Wed, 19 Aug 2026 00:00:00 GMTBUILDhomelabhttps://www.andraewilliams.com/posts/tool-use-giving-the-agent-hands/https://www.andraewilliams.com/posts/tool-use-giving-the-agent-hands/The tool schema is a contract, not a suggestion. Most tool-calling failures are the contract being too loose, not the model being too dumb.Wed, 19 Aug 2026 00:00:00 GMTTUTORIALaihttps://www.andraewilliams.com/posts/what-a-homelab-actually-is/https://www.andraewilliams.com/posts/what-a-homelab-actually-is/Not a datacenter in your closet. A place where breaking things is cheap — which changes what you should buy and what you should skip.Wed, 19 Aug 2026 00:00:00 GMTELI5homelabhttps://www.andraewilliams.com/posts/rag-explained-like-a-real-workflow/https://www.andraewilliams.com/posts/rag-explained-like-a-real-workflow/Retrieval-augmented generation is a research assistant with a filing cabinet. The failures are filing failures, not thinking failures.Tue, 18 Aug 2026 00:00:00 GMTELI5aihttps://www.andraewilliams.com/posts/what-mcp-actually-does/https://www.andraewilliams.com/posts/what-mcp-actually-does/The Model Context Protocol in plain terms — it is a port, not a plugin system, and that distinction explains most of the confusion.Tue, 18 Aug 2026 00:00:00 GMTTLDRaihttps://www.andraewilliams.com/posts/building-repotwin/https://www.andraewilliams.com/posts/building-repotwin/A repo intelligence twin — clone, index, embed, map, query. What worked, what I rewrote twice, and where it still falls down.Mon, 17 Aug 2026 00:00:00 GMTBUILDaihttps://www.andraewilliams.com/posts/running-small-local-models-on-my-homelab/https://www.andraewilliams.com/posts/running-small-local-models-on-my-homelab/Q6_K quantization, GGUF, and what 24GB of VRAM actually buys you in real throughput rather than benchmark numbers.Mon, 17 Aug 2026 00:00:00 GMTEXPERIMENThomelabaihttps://www.andraewilliams.com/posts/ai-agent-architecture-map/https://www.andraewilliams.com/posts/ai-agent-architecture-map/One diagram of how the pieces of a modern agent fit together — loop, tools, memory, evaluation, and the guardrails between them.Sun, 16 Aug 2026 00:00:00 GMTVISUALaihttps://www.andraewilliams.com/posts/how-i-structure-ai-agent-repos/https://www.andraewilliams.com/posts/how-i-structure-ai-agent-repos/A clean, boring layout that survives an agent growing from one script into a system — and the one rule that keeps it clean.Sun, 16 Aug 2026 00:00:00 GMTTUTORIALaihttps://www.andraewilliams.com/posts/cool-tools-for-ai-cyber-and-builders/https://www.andraewilliams.com/posts/cool-tools-for-ai-cyber-and-builders/The stack I actually use and recommend — not the one that makes a better screenshot.Sat, 15 Aug 2026 00:00:00 GMTRESOURCEaicyberhomelabhttps://www.andraewilliams.com/posts/fraud-controls-i-actually-look-for/https://www.andraewilliams.com/posts/fraud-controls-i-actually-look-for/The signals, rules and reviews that catch real fraud — and the theatrical ones that mostly generate work.Sat, 15 Aug 2026 00:00:00 GMTARCHITECTUREriskhttps://www.andraewilliams.com/posts/anime-interface-design-notes/https://www.andraewilliams.com/posts/anime-interface-design-notes/HUDs, typography and motion language from anime that hold up as real interface design — and the parts that only work on screen.Fri, 14 Aug 2026 00:00:00 GMTANIMEanimehttps://www.andraewilliams.com/posts/fight-iq-breakdown/https://www.andraewilliams.com/posts/fight-iq-breakdown/Angles, distance management and timing — the three things that separate a technical fighter from a fast one.Fri, 14 Aug 2026 00:00:00 GMTMMAmma