← back

A Little Behind the Scenes — Andrew Katana

Home lab notes, project logs, and occasional thoughts.

Why I'm Not Paying for ChatGPTNew

ChatGPT Plus, Claude Pro, Gemini — everyone starts with a subscription. I don't pay for any of them, and it's not a privacy screed. The real reason: I refuse to hand my thinking tools to someone else's pricing page. 85% that I control beats 100% that can change its pricing model tomorrow. The backlash-riddled Betamax argument: betting big on a model in flux is betting on a format war you don't get to vote in.

The Answer Was in Its Memory. The AI Fixed the Wrong Host Anyway.New

My AI had the one-line fix for a dead media server in its memory the entire time — then debugged the wrong host, blamed my clients, and cost forty minutes re-deriving what it already knew. Inside a shared hallucination: what happens when a confident agent and a busy human tell each other the same wrong story, and how to structure memory so it actually gets consulted.

My Linux Box Tunes Its Own TCPNew

A daemon on my dev box watches every TCP connection and quietly reconfigures the kernel to make it faster. 71,504 per-connection congestion-control decisions in one day, four algorithms matched to four kinds of paths, and the one operational catch: it forgets everything on reboot.

What Breaks First: People, Process, or Tech?New

Applying for jobs, I automate the process and fight the bots that check I'm not a bot — a human doing bot work so a bot can do human work. The same circus plays out across all of AI. The tech won't break; people adapt. It's the process that inverts, everywhere, in quiet absurd ways. A frame for the AI industry's real bottleneck.

My Pipeline Emailed Me Jobs I Already Applied ForNew

My job-search pipeline emailed me a daily digest of 28 ranked roles. Eight of them, I'd already applied to. The data existed on two machines — nobody had taught one about the other. The fix wasn't an LLM; it was a shared file, a cross-reference, and a loop we tested by breaking it. State is the product.

The Two-Week Rule of Agentic AINew

Two weeks ago my agents couldn't do what they do this morning. Agentic frameworks learn new tasks continuously — capabilities shift day to day, and what was impossible two weeks back is running in production today. Why documentation, evaluation, and the whole toolchain have a two-week shelf life, and the one approach that keeps working: play, push, discover.

The Week Everything Just WorkedNew

Last week my pipeline reported thirty successes and delivered zero usable results. This week it just worked — the first fully clean cycle in months. Here's what that actually cost, and why a clean week is compounding interest on debugging you did months ago.

What Nobody Tells You About Running Local LLMs: Seven Lessons From a Weekend of TuningNew

Selecting a local model is not a fire-and-forget process. Seven problems that no documentation page covers — from silently wrong context windows to GPU watchdogs to agents that fabricate their own actions. Practical lessons for anyone running local inference for real work.

The Pipeline Reported Success. Every Single Result Was Empty.New

My job pipeline printed a checkmark next to thirty results. Every single one was empty. No errors. No warnings. The system did exactly what it was built to do — and everything it produced was worthless. A field guide to silent failures in automation.

The Enterprise AI Router: Why Your Next Infrastructure Investment Isn't a Model, It's a BrokerNew

Every AI model you deploy creates a new attack surface — not from the model itself, but from the connection between the model and everything it touches. The missing layer in enterprise AI isn't better models. It's a broker.

The Post-Delivery Gap Nobody's Talking About

Demos work because they're small and simple. Production breaks when you introduce AI hallucinations and a loose approval process. Five war stories from two years of running AI agents in production — and three guardrails that actually work.

The Carwash Problem Has a Sequel — And It's WorseNew

The original Carwash Problem was visible. The sequel is silent. When AI gives you a confident, well-sourced, completely incomplete answer — because it can't join two separately-stated facts — that's the new failure mode nobody's testing for.

Why My AI Stack Lives on Three Machines (and Yours Should Too)New

Three ordinary machines that each do one thing well — a public-facing host, a GPU inference engine, and a development control plane. No Kubernetes, no service mesh, just three boxes speaking plain HTTP. Why distributed beats consolidated for home AI.

I Benchmarked Naive RAG vs. Graph-Based Retrieval on My Own Knowledge Base. The Results Were Not Close.New

I ran 20 queries against two retrieval architectures on my own 322-thought knowledge base. Naive RAG failed 65% of queries, mostly by refusing to answer questions the data could actually handle. A graph-based approach scored zero wrong. Here's the benchmark data, the failure analysis, and five practical fixes.

The Carwash Problem: Why Your IT Organization Isn't Ready for AI-Generated Code

When your AI agent kills the desktop it needs to operate, you learn fast that Enterprise IT's single-threaded pipeline can't absorb iterative, multi-source development. The Hyper-Loop framework and the three pillars of a safe operating environment for distributed production.

How My Homelab Sprawled to 12 Hosts and Why I'm Cutting It in Half

Twelve hosts, two hypervisors, a Raspberry Pi, and a managed switch. Most I chose. Some I accumulated. A practical guide to recognizing VM sprawl in your own homelab — and what happens when you actually write down everything on the network.

Self-Hosted AI Stack: Architecture Reference

The full inventory and network topology behind the self-hosted AI stack — 12 hosts, DMZ-vs-internal separation via CloudFlare tunnel, 10Gb data plane, MCP endpoints, and operational lessons from the migration that made it work.

Slop Coding and the Blind Spots AI Won't Tell You About

AI renamed my network interface. It asked for SSH to a box already sharing via NFS. It skipped 3TB of media without an error. These aren't edge cases — they're the current state. Three concepts and five war stories from running AI agents in the trenches.

ask.atkatana.com: Building a RAG-Powered AI from a SQLite Brain

How I built a production RAG pipeline using Open Brain (OB1), ChromaDB, and local LLMs — turning 322 career thoughts into a conversational AI with zero third-party API dependencies.

"AI Lost My Files" — And Other Lies I Told MyselfFeatured

A cautionary tale of AI-assisted data migration, ZFS backups that aren't rsyncs, and why the only bad actor was me. A lightly disguised position paper on managing access with vibe coding tools — born from the real loss of a Jellyfin library during a home lab migration.

I Ran My Job Search with Local AI Agents — Here's What Actually Worked

Three machines, five LLMs, zero API costs — how I built a local AI pipeline to automate my job search, and what it taught me about frameworks, hallucinations, and knowing when to write it yourself.

A Little Behind the Scenes: Building a Self-Hosted AI Hub — Andrew Katana

How and why this site is built on a home lab with local AI — a hands-on exercise in data sovereignty and independence.

Building a Cloud Center of Excellence from Zero

How I built a Cloud Center of Excellence from scratch at Micro Focus — 47 AWS accounts, Guardrails, IAM redesign, SRE transition, and the company's first legacy-to-Fargate SaaS migration.

The SaaS Maturity Model: How We Gave 10 Product Groups a Map to Cloud-Native

The framework that gave 10+ product groups a shared language for cloud adoption — six areas, progressive maturity levels, self-assessment methodology, and why the lab work mattered more than the slides.

From Data Center to ECS/Fargate in 12 Months

The technical deep-dive on migrating a heritage Java workload from over-provisioned data center VMs to AWS ECS Fargate — GitOps, Firecracker microVMs, 8-week migration timeline, and lessons learned.

The IOD Migration: 1,000 VMs, Four Failure Modes, and the Migration That Almost Didn't Happen

The migration that nearly broke the model — 1,000 VMs, eight years of organic growth, hand-operated monitoring, four unowned failure modes, and the organizational politics of cross-continent coordination.

The Local LLM Bet: Why Privacy Will Outlast the Hype

Privacy-first AI deployment, the VM sprawl parallel, and why local LLMs are a market — size depends on how the pay-per-play model executes over the next 3-5 years.

Two Truths and a Lie: Vibe Coding Edition

Two truths and a lie about vibe coding — the 80/20 trap, hallucinations in practice, and why you can actually do something about it.

The People Pipeline: Building SREs from Scratch

How we built SREs from scratch — career pipelines, retention math, compensation advocacy, and why a central CCoE is the right place to develop cloud talent.