AI Architecture

Specialized AI agents. One intelligent platform.

A multi-agent architecture that automatically selects the right AI model for every legal task — from contract drafting to compliance scanning.

GPT-4o-mini primaryClaude fallbackPerplexity for research

Architecture

How the AI Works

Four core systems power every AI interaction on the platform.

Model Orchestrator

Auto-routes tasks based on risk classification — formatting → nano model, legal reasoning → primary model, fallback → Claude.

Specialized AI Agents

CrewAI agents across 4 domains — contracts, compliance, vendors, analytics — each with distinct roles like Drafting Specialist and Risk Analyst.

Vector Intelligence

Semantic search across contracts, templates, and policies using 3072-dimensional embeddings powered by OpenAI text-embedding-3-large.

Real-Time Research

Perplexity Sonar API integration for live regulatory monitoring and vendor intelligence — not cached data, real-time research.

Models

AI Models in Production

Risk-based model selection ensures the right model handles each task.

ModelRoleWhen Used
GPT-4o-miniPrimary legal reasoningContracts, compliance, risk analysis
GPT-4.1-nanoLightweight tasksFormatting, classification, extraction
Claude 3.5 SonnetFallbackWhen primary model fails
Perplexity SonarLive researchRegulatory monitoring, vendor intel
text-embedding-3-largeEmbeddingsDocument search, template matching

AI Safety

AI Safety Pipeline

Every interaction passes through multiple safety layers before reaching AI models or returning to users.

Prompt Injection Detection

35+ injection patterns blocked before content reaches AI models.

PII Redaction

Emails, phone numbers, SSNs automatically stripped before external API calls.

Content Sanitization

All user inputs sanitized for XSS, SQLi, and path traversal patterns.

Response Validation

AI outputs validated for safety before being returned to users.

Input Length Limits

Configurable token limits prevent abuse and control API costs.

Audit Trail

Every AI interaction logged with model, tokens, duration, and user context.

Built for legal precision.

Specialized AI agents, risk-based model selection, and enterprise-grade safety — all working for your legal team.