Judge AI: OurFirm.ai's Judicial Intelligence Engine

Every trial attorney knows the feeling. The motion is solid, you know it in your gut, you've prepared for hours and the law is on your side. The only variable left is the one that decides everything: how this judge, in this courtroom, against this adversary, on these facts, will receive it.

Not what the law says in the abstract. What this judge does with it.

For as long as the profession has existed, that intelligence lived in hallway conversations, in war stories from colleagues who'd drawn the same judge, in unbillable hours buried in transcripts, and in a lot of educated guessing. The best litigators carried it in their gut. Everyone else filed and found out.

Judge AI ends the guessing. It builds a behavioral profile of your judge from the full record of their career. Not win/loss trivia, but how they think. How they handle your kind of motion. What they read closely and what they skim. Which authorities they reach for, how they structure their opinions, where a case like yours tends to turn, and the patterns that should shape how you brief and argue in front of them. Behind every profile sits a dataset of nearly 1.5 million judicial source documents. That dataset, more than anything else, is the story.

Litigation Reimagined, Not a Chatbot with a Law Degree

Most legal AI started life as a general-purpose chatbot. Vendors saw the legal market, bolted on the terminology, and called it a practice tool. You can feel it the first time you ask one a question only a litigator would ask.

OurFirm.ai started from the other direction, from the well of the courtroom. Before I wrote a word of product strategy, I spent fifteen years trying cases in state and federal court, and the hardest thing to capture in software is exactly what those years teach: what a judge's silence means, when to press an objection and when to sit down, which argument opens a door and which one closes the room. That's what we set out to encode. Judge AI is the clearest expression of that philosophy. Not a research add-on, not a summarization layer. Experience, structured.

I built OurFirm.ai because nothing on the market understood litigation the way litigators live it. Not as a chatbot, not as a search bar with a personality, but as a complete workflow: know the landscape, the players, and the judge; draft the motion; verify everything; anticipate the opposition; and walk into court ready to listen, push, and force the lean-in. Judge AI is the know-your-judge pillar of that workflow.

And it's built on a conviction about what AI is for. It isn't here to replace lawyers or outperform them. It's here to train, sharpen, and arm them. Judge AI doesn't argue the motion for you. It makes you the most prepared person in the room.

The first release delivers deeply drawn behavioral profiles of the federal bench. Not just how a judge has ruled, but why. How they reason, their temperament, their tone, what makes them lean in, and what actually moves them.

What Judge AI Does

Judge AI runs a judge's entire record through our proprietary behavioral extraction pipeline: published and unpublished opinions, hearing transcripts, orders, and docket entries, not just the decisions that make it into a case reporter. The pipeline pulls exact quotes and decisional signals straight from the judge's own words, and returns them with Bluebook citations linked to the source document. You don't get a guess. You get the judge, in the judge's own words, and you can verify every quote yourself.

The pipeline also captures the legal standards a judge reaches for and the analytical logic that runs through their decisions. That layer lets you do more than read about your judge. It lets you mirror your judge, and turn a generic brief into an argument built for the one person deciding it.

Here is what that makes possible:

• See how your judge treats motions like yours before you draft a single page.
• Match your argument structure and authorities to the ones this judge actually responds to.
• Walk into oral argument already knowing the questions waiting for you at the bench.
• Echo your judge's own language and reasoning back to them in your brief, because judges respond to their own words.
• Cite your judge's prior rulings, published and unpublished, back to them.
• Calibrate how hard to press an objection based on how this judge has ruled on it before.

You pressure-test your theory of the case against how your judge has actually ruled, before you ever step into a courtroom.

None of it works without the dataset underneath. And the dataset is the part nobody else can copy.

Why Data Is the Moat

Every legal AI product on the market is built on the same handful of frontier models, and those models improve every quarter, for everyone, simultaneously. Whatever intelligence we can rent, our competitors can rent. So the model can't be the moat. The model is for sale. The only durable advantage is what your platform knows that the model doesn't.

Here's what no foundation model knows: your judge's actual record. Models are trained on the public web, and on the public web a judge is mostly their published opinions, the small, curated fraction of their output that reaches a reporter. No reporter publishes temperament. No headnote captures how a judge runs a status conference. The record that actually predicts judicial behavior lives everywhere else, in the minute orders, unpublished rulings, hearing transcripts, and docket entries where litigation gets decided day to day. And that record is technically public and practically dark: scattered across PACER, unindexed, unstructured, priced by the page. Fragments surface online, but a model's diffuse memory of a transcript is not a dataset. You can't query it by judge and issue, you can't check its work, and you can't cite it in a brief.

So we built the dataset. And because in this profession the counting method matters as much as the count, every number below comes from the qualified, audit-complete layer. These are the figures that survived verification, not the totals we downloaded:

• 1,118 federal judges and judicial officers, including 788 U.S. District Judges and 280 U.S. Magistrate Judges, across 82 federal court designations.
• 1,495,654 judicial source records: published and unpublished opinions, hearing transcripts, orders, and docket entries.
• 9,824,323 structured intelligence records: decision, reasoning, behavioral, and language records built from those sources.
• 3,351,992 recovered decision units, including rulings that never reached a reporter and never will surface in a tool built on the published record.
• 687,543 prediction-ready decision units: a focused evidentiary layer supporting judge-specific research, retrieval, and analytical workflows.
• 784,348 issue-level tags: every decision organized by evidentiary issue, procedural posture, outcome, and reasoning characteristics.
• 1,498,466 judicial-behavior observations: courtroom management, oral-argument dynamics, guidance from the bench; more than 1,300 per judge, on average.
• 1,405,457 language and style signals: the recurring phrases and rhetorical moves that let you echo a judge's own words back to them.
• 13,859 judge-specific reasoning patterns, grounded in 63,864 supporting decision-unit links, with a reasoning profile for 99.6% of the bench we cover.

Read those numbers top to bottom and you're looking at the moat itself. This isn't a scrape; it's a refinery. Source documents become structured records, structured records yield recovered rulings, rulings get tagged by issue and posture, the behavioral and language layers sit on top of that, and the reasoning patterns at the summit are grounded, link by link, in the decision units below them. Each layer is derived from the one beneath it, which means there's no shortcut to the top. A competitor can't generate the reasoning patterns without recovering the decision units, and can't recover the decision units without first assembling and structuring the record. Starting today, they start a million and a half documents behind. And the corpus grows every docket day, because every new opinion, order, and transcript runs through the same pipeline. The gap doesn't close. It compounds.

Scale is the easy part to describe. Judgment is the hard part to copy. A transcript is just paper until someone knows that a judge's aside at a status conference is signal and a boilerplate scheduling order is noise. The signals our pipeline extracts weren't chosen by engineers guessing at what lawyers might want. They encode what fifteen years of standing in front of judges taught me actually moves a courtroom. The schema is trial experience written into extraction logic, and it doesn't ship in anyone's API.

And the final layer is the one this profession is rightly unforgiving about: every record traces back to a verifiable, Bluebook-cited source document. In law, an insight you can't cite is an insight you can't use. Analytics you can't verify are a horoscope. Clearing that bar at nine-million-record scale is expensive, which is exactly why the figures above say qualified instead of collected. We publish what survived the audit.

One more thing competitors have backwards: every frontier model release makes this dataset more valuable, not less. A better model reasoning over 9.8 million structured records beats the same model reasoning over headnotes, and that delta widens with every release.

The data is the moat. The agent architecture we engineered for litigation workflows is the second. The outcomes attorneys are seeing in live cases are the third.

See It on Your Judge

Don't take the numbers on faith. Bring us a pending matter and we'll show you the profile we build for your bench: how your judge has handled issues like yours, in their own words, with the citations to prove it.