All case studies
Legal / FinTech · Lexpertz AI

Axiom Engine V4: zero-hallucination document auditing

Engineered a deterministic multi-agent LangGraph circuit with an adversarial prosecutor node. Uses MCP Protocol to bridge local IDEs to a FastAPI cloud backend. RAGAS Faithfulness drops trigger cyclical retries — achieving 99.9% verifiable output on SEC filings and legal MSAs.

Citation accuracy99.9%
Container savings2.5GB
Cold boot time<1s

The brief

Legal and finance review teams do not need a faster summariser. They need a system that can show which sentence in which filing supports which claim, and that refuses when the evidence is not there.

What we built

A deterministic LangGraph circuit with an adversarial prosecutor node: one agent assembles the answer, a second attacks it against the retrieved passages. Retrieval is mediated by MCP so local IDE tools and the FastAPI backend share one tool contract. When RAGAS faithfulness drops below the gate, the circuit retries with a narrower retrieval window instead of guessing.

What it proved

99.9% verifiable output across SEC filings and legal MSAs in the evaluation set, with the runtime container down 2.5GB and cold boot under a second. The failure mode that remains is a missing document — which the system reports rather than hides.

  • RAG
  • LangGraph
  • MCP
  • FastAPI
  • pgvector