A court case is a data problem nobody treats like one
A mid-sized dispute is thousands of pages: pleadings, e-mails, scanned letters, WhatsApp exports, recordings, court minutes. They arrive out of order, in different formats, over months. The witness who said one thing in March says another in October, on page 1,412 of a PDF nobody re-reads.
Lawyers handle this with highlighters, Word tables and memory. I have built software for 25 years, and I kept looking at those files thinking: this is ingestion, extraction, indexing and diffing. It is a data pipeline with very high stakes and very low tolerance for hallucination.
So in May 2026 a team of four of us started building Legalnaut. It launched in October. This post is about what it does and the engineering decisions that mattered most.
What Legalnaut does
You drop in a case file, often as one zip, 7z or rar archive. Legalnaut reads it once and builds a working model of the case:
- Chronology. Dates, people and events from every document on one timeline.
- Contradictions. Conflicting passages shown side by side. Nothing counts as a contradiction until a lawyer confirms it.
- Proof chart. Each issue matched to its evidence, with the gaps visible. Chronologies and bundles export with exhibit numbers.
- Hearing preparation. Witness dossiers, examination outlines and deadlines calculated from the rule and the start date.
- AI assistant. Searches and reads the file, scans and recordings included, and answers with links to exact pages.
- Client access. Clients see their documents, chronology, hearings and deadlines, and get a daily e-mail when something new is filed.
Inputs are messy on purpose: PDF, Office and OpenDocument files, Outlook mailboxes, images, audio and video. Scans go through OCR, recordings through transcription. Every matter also gets its own mailbox address, so documents can be forwarded straight into the file.
What it does not do matters just as much. It gives no legal advice, files nothing with a court and does no e-discovery privilege review. Its findings are suggestions until a lawyer signs off.
Six engineering decisions that shaped it
1. No claim without a source passage
Every extracted claim, event, person and organisation is stored with a pointer to the passage it came from. The UI never shows a fact it cannot link to a page.
This sounds like a UX nicety. It is actually the core data model. If a record has no source span, it does not exist.
2. Flag, never repair
Citations in AI answers and drafts are checked against the file after generation. When a citation does not hold up, the obvious move is to ask the model to fix it. We deliberately don't.
A "fixed" citation is a second guess dressed up as a fact. In a courtroom that is worse than no answer. So unverifiable claims are flagged and left visibly broken for a human to judge.
3. The model proposes, a lawyer decides
Contradiction detection is the feature people get excited about, and the one most likely to embarrass you. Two passages that look contradictory often aren't once you know the context.
So a contradiction is a suggestion with two linked passages until a lawyer confirms it. Confirmed and unconfirmed findings are different states in the data, not just different badges.
4. Evidence is append-only
Once a document enters a matter, its file, extracted text and fingerprint cannot be overwritten. Files are identified by SHA-256, so the same e-mail uploaded three times is one document, analysed and billed once.
Immutability made later features cheap. Exhibit numbers stay stable, audit logs mean something, and "which version did we rely on" has one answer.
5. One database per matter
Each matter lives in its own isolated database, and search cannot cross matters. A law firm holds files for clients who may be on opposite sides of other disputes. A missing WHERE clause should not be able to leak one client's evidence into another's answer.
It costs more operationally than row-level tenancy. For this domain I think it is the only defensible choice. Hosting is in Germany, and customer content is not used to train models.
6. Bill in pages, not tokens
Lawyers think in pages, not tokens. So the only metered unit is a page: one PDF page, 400 words of e-mail or text, or one minute of audio.
The rules behind it were more work than the pricing page suggests:
- A page is counted once, when the file is first analysed. Re-reading, chat, search and exports are free.
- Duplicates are free, thanks to the SHA-256 check above.
- Files with no readable text are refunded.
- An estimate is shown before an import starts.
- When the allowance runs out, new files wait in a queue instead of being rejected.
The last one is my favourite. Nothing a user uploads is ever lost because of a billing state. The same principle applies to failed payments: the account goes read-only, but reading, search and export stay open, and nothing is deleted automatically.
Try it on a real case
The fastest way to see it is the public demo. No sign-up, read-only, loaded with the court filings from Epic Games v. Apple. Ask the assistant something and click the citations; that is the whole pitch.
To run your own documents, there is a 7-day trial with no card: one matter and 1,000 pages. Paid plans start at $79 a month, with unlimited users on every plan (pricing).
This week only: 30% off any plan for 12 months with code BF2026, valid until 30 November.
If you build with LLMs in a domain where being wrong is expensive, I would like to hear how you handle grounding and verification. Drop a comment, or tell me where you think the "flag, never repair" rule breaks down.
Top comments (0)