Startseite AI Legal Assistant

AI Legal Assistant

Headline From trial notes to cited case law. In a minute., next to a laptop showing the draft assistant and a card with a cited paragraph from a European Court of Human Rights judgment

Services:

  • Product discovery
  • UX/UI design
  • Full-stack development
  • AI engineering
  • Data pipeline & deployment

Industry:

  • Human rights · Legal tech · Non-profit

Technology:

  • Angular 21
  • NestJS / TypeScript
  • MariaDB
  • OpenAI
  • LangChain
  • Google Sign-In
  • Google Sheets API

Challenge

  • „Every violation I allege has to rest on a paragraph where the Court states a principle. Finding it is an afternoon in HUDOC.“
  • „Our senior lawyers already know the best paragraphs for hundreds of cases. They are in a spreadsheet only they can navigate.“
  • „Half the team writes in Ukrainian. The judgments are in English and French.“
  • „An AI that invents a citation is worse than no AI. Every suggestion has to be something a lawyer would sign.“

Results

  • Candidate violations and citable paragraphs in about a minute per report; drafting time roughly halved by the team’s estimate (exact hours to confirm)
  • 87–89% agreement with the lawyers‘ own paragraph picks in blind replays of real drafts, so the first draft already reads like a senior colleague’s
  • 200+ judgments and 27,000+ paragraphs in a reviewed knowledge base; every cited paragraph identical to the Court’s text, so nothing pasted into a report needs re-checking
  • Ukrainian notes retrieve French judgments as reliably as English ones, with no translation layer
  • Live with real observers; every like and dislike is stored against the exact draft text and feeds the next tuning round

Good fit for

  • Legal, compliance and research teams that must cite the source, not summarize it
  • NGOs and monitoring missions publishing evidence-based reports on deadlines
  • Any organization whose best knowledge sits in one expert’s spreadsheet
  • Distributed teams working across languages

You leave the courtroom with three pages of notes and a report due tomorrow. Somewhere in 27,000 paragraphs of European Court of Human Rights case law is the sentence that turns what you saw into an allegation a judge will take seriously. You paste your notes. About a minute later the assistant names the violations they point to and hands you the paragraphs to cite, each one verified word for word against the Court’s own database.

Infographic: six steps from pasted trial notes to cited paragraphs, with the Smart Search progress dialog and the line Never silently dropped

One paste, about a minute, and nothing invented. The assistant reads your notes the way your lawyers would. It names the violations under the organization’s own 20-topic taxonomy, splits the text into separate complaints, searches each one semantically and re-ranks what comes back. Then it asks the only question that matters: could a lawyer honestly cite this paragraph for this complaint? Anything that fails is set aside where you can see it, never silently dropped. Each suggestion arrives with the full paragraph, the case, the date and a one-line reason in your language. The team estimates report drafting time has roughly halved.

Infographic on indigo: lawyers' spreadsheet, scheduled sync and HUDOC feeding the Cases table in a browser card, with the figures 200+ judgments, 27,000+ paragraphs, 100% byte-identical

A reviewed knowledge base, not a web crawl. The senior lawyers‘ spreadsheet became the ground truth. A scheduled sync reads their reviewed labels, fetches each judgment from HUDOC, picks the right version when one application number has several (Chamber, Grand Chamber, English and French originals) and stores exactly the paragraphs they cited. The corpus holds 200+ judgments and 27,000+ paragraphs, and every cited paragraph is byte-identical to the Court’s published text. In blind replays of real drafts the assistant agreed with the lawyers‘ own picks 87–89% of the time.

Notes in Ukrainian, English and French pointing to the same verified paragraph card from Morabito v. Italy

Any language in, the right judgment out. Observers write in Ukrainian; the Court writes in English and French. Multilingual embeddings find the same paragraph whatever language it was written in, so there is no translation step to introduce errors. Three interface languages, Google sign-in for observers, admin-only access to the corpus and taxonomy tools, workspaces that follow you between devices, and a documented REST API so partner teams can plug the same search into their own agents.

If this sounds like your team: expert knowledge a few people carry in their heads or their spreadsheets, and everyone else waits for them. We turn it into a tool the whole team can trust.

Lass uns anfangen!











    Diese Website ist durch reCAPTCHA geschützt und unterliegt der Anwendung von
    Datenschutzrichtlinien та
    Nutzungsbedingungen від Google.


    Wunderbar!

    Wir kommen gleich zurück, um ordentlich zu plaudern.

    Kostenlose Beratung