legal-due-diligence-orchestrator

An Agentic AI system that operates as an autonomous legal due diligence orchestrator. Unlike linear automations, this multi-agent architecture utilizes an adversarial reasoning loop to audit complex contracts, make independent decisions about compliance risks, and use tools to securely parse documents, verify absence claims, and generate deterministic risk reports.

Instead of manually skimming hundreds of pages to prove what an agreement doesn't say, this system gives AI agents access to your legal documents and compliance rulebooks, allowing them to execute your firm's Standard Operating Procedures (SOPs) autonomously while actively preventing "lazy LLM" false negatives.

Contract review tools are good at telling you what a document says. The expensive paralegal work is proving what a document doesn't say — that there's no cap on damages, no governing-law clause, no return-of-materials obligation.

That's also where language models fail hardest, and fail quietly. A model that skims a 90-page master services agreement and reports "no limitation of liability found" produces output identical to a model that looked carefully and found nothing. Nothing in the response distinguishes diligence from laziness — and in legal work that false negative is the difference between a clean deal and uncapped exposure.

So this system treats absence claims as untrusted by default. Any agent that reports a missing clause triggers a verifier that re-reads the document with a deliberately different retrieval strategy, on a different model, with the burden of proof inverted — before the finding is allowed into the report.

[ Ingestion ]  pypdf → paragraphs anchored to (page, ¶, §)
      │
      ▼
[ Lead Agent ]  loads SOP, scopes each rule to relevant sections
      │
      ├──► [ Liability Worker ]     ──► "uncapped risk"
      ├──► [ Termination Worker ]   ──► "clause missing"
      ├──► [ Jurisdiction Worker ]  ──► "found, §8"
      │
      ▼
[ Verifier ]  synonym search → section scan → full text
      │        goal: PROVE THE WORKER WRONG
      ├──► one verifiable hit ⇒ finding overturned
      │
      ▼
[ Output Agent ]  Pydantic-validated JSON + risk score

• The "Lazy LLM" Problem: A model skimming a 90-page MSA will often say "no clause found" just to save compute.
  • Solution: Adversarial Verification. The verifier is forced to use a different retrieval strategy (synonym search + section scans) and a different model (Gemini vs. Claude), with the burden of proof inverted.
    
    
• Hallucinated Quotes: LLMs notoriously invent text that sounds legally plausible.
  • Solution: Citation Gates. A custom cite_source() function verifies quotes against the anchored (page, ¶, §) source text before minting a Citation object.
    
    
• Self-Contradictory JSON: Agents often output a status of "MISSING" while providing a quote of the clause.
  • Solution: Schema-Enforced Integrity. Pydantic v2 with extra="forbid" throws a ValidationError if an agent submits logically impossible combinations.