How Law-Bunker works

Built on Fireworks

Law-Bunker is a lawyer and a set of AI models working together. The machines find the rulings and prepare the groundwork; I read, explain and record every story myself. Every model in the pipeline runs on Fireworks AI, with one exception, explained below.

How it works

The work moves through seven boxes. Each box has one job, takes in one fixed thing and hands on one fixed thing, so any box can be fixed, or have its model swapped, without disturbing the others.

Why one box isn't on Fireworks: Fireworks has retired audio inference, so speech-to-text runs on OpenAI's open Whisper large-v3 through OpenRouter, with zero data retention required on every call. Everything else that uses a model runs on Fireworks.

How I use it

Preparation runs overnight. In the morning there are three prepared cases waiting, each with its outline next to the ruling, and a short note on my phone with what's ready and what it cost.

1. Pick. I read the outlines beside the rulings and choose one or two worth telling.
2. Draft. I record a draft from the outline, the way I'd explain it to a friend, into a synced folder. A small home server picks it up and matches it to the case by its name.
3. Transcribe. Speech-to-text turns it into text. Filler — "um", false starts, repeated words — is removed by fixed rules, never by a model, and asides that don't belong in the script are lifted out and shown to me separately.
4. Polish. I choose a full polish in my style, or a light clean-up that only fixes slips and keeps my order and my jokes. The changes are computed in code, one sentence at a time, and I accept or reject each one.
5. Check. A model from a different family reads my script against the ruling and flags anything that doesn't match — a wrong court name, say, or a holding stated backwards. It never rewrites; I decide on each flag.
6. Record the final. I re-record the finished script. Its transcript gets one last check, and only then can it be published.

How it improves my output

After every recording, Learn compares the script I was given with what I actually said, sentence by sentence, and turns the differences into concrete style lessons with evidence. Those lessons shape the next script — but only the ones I approve.

On the first real case, McCain v. Mary Mahoney's, the machine gave me 817 words and I said 821. Of its 43 sentences:

A few of the lessons it drew:

• No signposting. I cut every "That concludes Act One" and wove the transitions into the story instead.
• Short punchlines. "Seafood, fraud, and 29 tons of fish with falsified passports" survived; the wind-up before it did not.
• Plain-English glosses. "Denied him leave to amend as futile" became "did not give him permission to amend his complaint to change these allegations and perhaps continue his lawsuit."
• Name the court every time. For listeners, "the District Court said" beats "the court" with context they can't scroll back to.

Building it

I don't program. Law-Bunker was built as a working partnership: I set the rules and made every editorial decision, and an engineer friend and AI coding assistants turned them into the pipeline. A few principles held throughout:

• Boxes with contracts. Each of the seven boxes was built and tested alone against a fixed hand-off before the next was joined to it. A bad hand-off is rejected loudly.
• A stand-in for the lawyer. While it was being built, a small, cheap model played my part at every step marked "me", so the plumbing could be tested end to end without me. Its output is marked, never published, and never learned from.
• Simulated days. Find, Collect and Prepare were run over past dates, using only information available on each day, before anything ran live.
• Everything on the record. Every model call keeps its prompt, model, settings and cost. Any news article used to find a story is archived privately with a timestamp, as the fair-use record; its text never appears here.
• Honest failures. When a box can't do its job, it says DEGRADED rather than reporting zeros as if all were well.
• Hard spending caps on every paid service, checked before each call.
• Fireworks' own guidance as the standard — its recommended models for each kind of job, prompt caching (the ruling goes first, word for word, in every call about a case), and no features that store prompts or outputs.

What it costs to run

Running costs only, measured from the pipeline's own spending ledger as of 19 October 2026. The website itself runs on Cloudflare's free tier, and court records come from the free tier of CourtListener.

One real story, start to finish

Projected monthly operating cost, by box

At the planned pace — three rulings prepared every day, one or two of them voiced (1.5 on average), and Learn running once a day — Law-Bunker costs about $0.81 a day, or $24 a month.

Rulings that have to be bought from PACER are extra, and are bought by hand only when a story is worth it.

On the horizon

What's planned next. None of it exists yet.

And two changes behind the scenes:

• Automatic audio cleanup. Every final recording cleaned before it's posted — background noise and room echo removed, loudness evened out to podcast standard, and filler sounds cut using the transcript's word timings — on the home server, so no audio leaves the house.
• Fine-tuning. A long-term look at whether my accepted and rejected edits can fine-tune a model on Fireworks that writes the first draft in my voice. See The Law-Bunker model.