Applied AI02

CRIS

Contract Risk Intelligence System

A multi-agent system that audits engineering contracts for risk exhaustively, not merely until a reviewer reaches the end of the document.

coordinated agents
validation layers
confidence threshold for human review
Multi-agent RAGVector retrievalConfidence thresholds
ClaudePineconen8nPostgreSQL
Também usado
Private code - demonstration available on request

CRIS is a multi-agent contract-auditing system built with Claude and eleven specialized agents, each with its own retrieval layer. It expands a legal team’s ability to surface contractual risk while keeping the final decision with people.

At a glance

ProblemHigh-value contracts contain interconnected risks that manual review cannot exhaustively map.
SolutionSpecialized agents with retrieval and three layers of validation scaled to contract value.
ResultMore than R$5M in contractual exposure mapped, with explicit uncertainty below the review threshold.

The problem

Engineering contracts can combine hundreds of pages of technical and legal clauses with exposure in the tens of millions of reais. Important obligations are often hidden in how clauses interact.

A manual review can take weeks and still leave uncertainty about what was missed. The system was designed to extend, not replace, the legal team’s review.

Impact

Risk review becomes systematic rather than sample-based, giving decision-makers traceable findings before and during execution.

The project has mapped more than R$5M in contractual exposure in real use. This is mapped exposure, not loss incurred.

Human reviewers retain the final decision, especially when the system’s confidence falls below its threshold.

Confidence is not binary. Making uncertainty explicit is what makes an AI system useful for high-stakes contracts.

Architecture

Claude coordinates eleven specialized agents. Each retrieves from the full contract before returning a finding with evidence and confidence.

Diagrama da arquitetura do CRIS
Fig. 01 — Arquitetura do pipeline
IngestionOCR and structure-aware chunking
Specialized agentsEleven Claude agents with dedicated retrieval
ResilienceSupabase checkpoints for resumable processing
ValidationThree checks scaled to contract value