Pacific Design/ artificial intelligence

about · who writes this, and why it is free

Who writes this.

Kevin Thomas Luzbetak, MSc — an AI/ML engineer in Los Angeles. I build retrieval systems, agents and document-intelligence pipelines for people who have to be able to check the answer. This site is where I write down what that work turns up.

Head-and-shoulders photograph of Kevin Thomas Luzbetak: short dark hair, thin-rimmed glasses, looking straight at the camera against a blurred outdoor background.
Kevin Thomas Luzbetak, MSc Los Angeles, California

Pacific Design is not a company, a team or a content operation. One person writes the entries, argues with them, builds the sixteen instruments in the Lab and runs the machine they arrive from. Where the home page says "rather than take our word for them" — that our is me.

The day job

I build document intelligence: pipelines that take a customer's PDFs — contracts, clinical paperwork, discovery — and answer questions about them with the page and the paragraph attached. The current one extracts clauses across forty-plus named types, scores them against a configurable playbook, and writes one immutable audit record per answer, hashed, so a customer can recompute it and prove nothing was edited after the fact.

What took longest to learn is what this site keeps coming back to: off-the-shelf retrieval does not survive a contract. Fixed-size chunking cuts a numbered section in half; dense vectors alone miss the clause with the case number in it. What works is keyword and embedding search fused by rank, a cross-encoder over the survivors, and a router that sends anything carrying patient data to a local model that cannot reach a frontier API over the network at all. I took a Master of Science in artificial intelligence in 2014, which was a quieter year for it.

Why the map is shaped this way

All fifteen sections are written with the same bias: toward the decisions that arrive with a bill attached. Agents spends its room on loops, tool calls and structured output, because a model that returns prose your program cannot parse is an outage. RAG & Embeddings comes early in the reading order because retrieval is where these products actually fail. LLMs and Infrastructure carry routing, serving and what a GPU hour really buys. Applied & Classical ML is here because most machine learning in production is still rows and columns, and the field stopped writing about it. Every entry ends on the way the thing fails, which is the part you need once it already has.

Why it is free, and what the server knows

Because serving it costs me almost nothing and the work was done anyway. Eight sites carry my name, the oldest two registered in 1997. Five of them, this one included, are static HTML served straight off disk — 14,177 pages between them, no database on the request path, search built offline. Static by default; dynamic only where a feature genuinely cannot work otherwise. That is not nostalgia, it is what lets one person run eight production sites without a platform team.

For you, that same rule is the privacy policy. No reader account, because there is nothing for you to log into. No cookies, no trackers, no third-party scripts, no fonts or analytics fetched from anyone else: the Lab cannot phone home because there is nothing on the other end, and the glossary and index filter inside your browser without telling the server what you typed. What the server does keep is an ordinary Apache access log — your address, the page, the referrer — which I read in aggregate to see which entries get read. That is the whole of it, and I would rather write it down than let "nothing tracked" quietly mean more than it can.

The eight

Oldest domain first; the years are when the domain was registered, not when the site was any good. Same rule throughout: I design, build and operate them end to end, from the Apache vhost up.

Corrections

A free site earns trust by being checked, including against itself. Everything above is my own account of my own work, and nobody edits me but the arithmetic. If an entry here is wrong, out of date, or sounds confident where the evidence is thin, say which part and I will fix it or say why not.

New here instead? The six reading paths will pick a route for you.