Multi-Agent Litigation AI

LexOrchestrator

Eight specialist agents running sequentially — retrieval, drafting, citation verification, adversarial review, local rules, judge intelligence, and eval scoring — on every litigation workflow.

Start a Draft

The Problem

A single LLM call produces plausible text with no guarantee of accuracy.

Hallucinated citations

A bare LLM call invents case names, reporters, and page numbers that don't exist. No model self-corrects without a retrieval loop.

No adversarial pressure

A single-pass draft doesn't account for opposing counsel's strongest counterarguments. Motions that aren't stress-tested get shredded at hearing.

Local rules blind spots

Formatting, page limits, and required-section rules vary by district and judge. One missed requirement and the clerk rejects the filing.

Specialist Agents

Sequential pipeline

Hybrid RAG

Retrieval Method

pgvector + keyword

MCP Tools

Claude tool integration

Full Trace

Observability

Every step inspectable

The Pipeline

Eight agents. Sequential execution.
Every step inspectable in real time.

Each agent receives the accumulated context from all prior agents. Events are streamed to the trace log as they occur.

Intake

Retrieval

Drafting

Citation

Adversarial

LocalRules

JudgeBrief

Eval

IntakeAgent

Classification

Normalises the request: detects motion type, jurisdiction, court, key legal issues, and missing inputs. No LLM call wasted on a malformed request.

Raw query + facts

Out

Intake record

RetrievalAgent

Hybrid RAG

Fires parallel pgvector cosine search and BM25 keyword retrieval over indexed court opinions. Fuses results with a weighted scoring formula, boosting primary sources and jurisdictional matches.

Legal issues + jurisdiction

Out

Ranked opinion chunks

DraftingAgent

Generation

Produces a fully-structured motion draft grounded in retrieved authority — Preliminary Statement, Statement of Facts, Legal Standard, Argument (Roman numeral sections), and Conclusion.

Ranked chunks + facts

Out

Cited motion draft

CitationAgent

Verification

Extracts every citation using regex + optional eyecite worker. Verifies each against the corpus: existence, reporter format, pin cite page, quote fidelity, and proposition support.

Draft text

Out

Citation report

AdversarialAgent

Red-Team

Simulates opposing counsel. Identifies the three strongest counterarguments, flags unsupported claims, assesses adversarial risk (HIGH / MEDIUM / LOW), and suggests rebuttal paths.

Draft + citations

Out

Red-team memo

LocalRulesAgent

Compliance

Checks the draft section-by-section against jurisdiction-specific local rules profiles (SDNY, Federal, New York State). Reports missing sections, formatting violations, and filing notes.

Draft + jurisdiction

Out

Rules report

JudgeBriefAgent

Intelligence

Queries cached judge profiles derived from prior opinions. Returns citation preferences, writing-style notes, motion-type-specific guidance, and risk flags for the assigned judge.

Judge name / ID

Out

Judge brief

EvalAgent

Scoring

Computes a weighted confidence score across six dimensions. Produces a pass / warn / fail verdict and a structured quality report persisted alongside the draft.

All agent outputs

Out

Confidence score

Sample Output

What the pipeline actually produces.

Motion to Dismiss — S.D.N.Y. — Demo RunCompleted82% confidence

PRELIMINARY STATEMENT

Defendant Northstar Retail Systems respectfully moves to dismiss
pursuant to Federal Rule of Civil Procedure 12(b)(6). Plaintiff's
claims fail as a matter of law: the Pilot Agreement expressly
conditioned any production deployment on a later signed order form.
No order form was executed.

LEGAL STANDARD

To survive a motion to dismiss, a complaint must plead "enough facts
to state a claim to relief that is plausible on its face." Bell
Atlantic Corp. v. Twombly, 550 U.S. 544, 570 (2007). A claim is
plausible "when the plaintiff pleads factual content that allows the
court to draw the reasonable inference that the defendant is liable."
Ashcroft v. Iqbal, 556 U.S. 662, 678 (2009).

ARGUMENT

I.  THE BREACH OF CONTRACT CLAIM FAILS

    Where sophisticated parties have negotiated in anticipation of a
    formal written agreement, no binding contract arises until that
    document is executed. R.G. Group, Inc. v. Horn & Hardart Co.,
    751 F.2d 69, 75 (2d Cir. 1984). The Pilot Agreement's plain
    language forecloses any contrary inference.

[  CitationAgent: 5/6 verified  ·  AdversarialRisk: MEDIUM  ·  Confidence: 82%  ]

View full demo →

Under the Hood

Deterministic scoring formulas,
not black-box outputs.

Retrieval Scoring

score = min(1,
  kwScore    × 0.35
+ vecScore   × 0.45
+ jurisdictionBoost
+ courtBoost
+ authorityBoost   // +0.15 for primary sources
+ recencyBoost
)

Hybrid fusion of pgvector cosine similarity and BM25-style keyword overlap. Primary sources (Constitution, statute) receive an authority boost.

Confidence Scoring

Citation pass rate35%

Faithfulness score25%

Retrieval coverage15%

Local rules completeness10%

Adversarial safety10%

Judge alignment5%

Pass ≥ 75% · Warn ≥ 55% · Fail below. Internal quality signal only.

Tech Stack

Next.js 16

App Router + Turbopack

TypeScript

Strict mode, end-to-end

Supabase

Postgres + pgvector

OpenAI / OR

LLM + embeddings

eyecite

Legal citation extraction

pdfkit + docx

Server-side export

MCP stdio

Claude tool integration

Tailwind CSS

Design system

See the pipeline in action.

Run a motion draft end-to-end in under 30 seconds.

Start a Draft

Demo corpus only. Not real legal authority. Not legal advice. LexOrchestrator is a portfolio project demonstrating litigation AI architecture.