Skip to content

Fulcrum System Overview

Executive Summary

Fulcrum is an enterprise-grade cognitive governance platform that provides runtime safety, policy enforcement, and cost management for AI agent deployments. Unlike static rule engines, Fulcrum's cognitive architecture understands intent, predicts failures, and evolves defenses automatically.

Core Value Proposition

The Problem: Enterprises deploying AI agents face existential risks: - Runaway agents consuming unlimited resources - Agents taking unauthorized actions - Compliance violations from uncontrolled behavior - No visibility into agent decision-making

The Solution: Fulcrum provides "operational insurance" through: - Real-time governance (<10ms policy evaluation) - Intent understanding (Semantic Judge cognitive analysis) - Predictive cost control (Oracle budget prediction) - Adaptive defense (Immune System policy generation)

Architecture Layers

Layer 1: Execution Envelope

The execution envelope wraps every AI agent action: - State machine tracking (PENDING → RUNNING → COMPLETED/FAILED) - Immutable event logging - Cost accumulation - Policy checkpoint integration

Layer 2: Policy Engine

Deterministic rule evaluation: - JSON-based policy definitions - <10ms P99 latency - Redis-cached evaluations - Tenant-scoped policies

Layer 3: Cognitive Layer (The Brain)

AI-powered governance components:

Component Function Latency Target
Semantic Judge Intent analysis via LLM <50ms
Oracle Cost prediction <20ms
Immune System Auto-policy generation Async

Layer 4: Event Store

Durable event persistence: - NATS JetStream for real-time streaming - PostgreSQL + TimescaleDB for analytics - Full audit trail with RLS isolation

Data Flow

Agent Request
    ↓
[Execution Envelope Created]
    ↓
[Policy Engine Evaluation] ←→ [Redis Cache]
    ↓
[Cognitive Analysis] ←→ [Ollama LLM]
    ↓
Decision: ALLOW | DENY | REQUIRE_APPROVAL
    ↓
[Event Published] → [NATS JetStream]
    ↓
[Dashboard Display] ←→ [PostgreSQL]

Technology Stack

Component Technology Rationale
Backend Go 1.24 Performance, type safety
Database PostgreSQL 16 + TimescaleDB ACID, time-series
Cache Redis 7 Sub-ms policy lookups
Message Queue NATS JetStream Durable streaming
LLM Inference Ollama Local, private, fast
Frontend Next.js 14 React ecosystem
Auth Clerk Enterprise-ready
Protocols gRPC + Protobuf Type-safe contracts

Multi-Tenancy Model

Fulcrum implements strict tenant isolation:

  1. Database Level: PostgreSQL Row Level Security (RLS)
  2. Application Level: Tenant context injection
  3. Cache Level: Namespaced Redis keys
  4. Event Level: Tenant-scoped NATS subjects

Every query automatically filters by tenant: 

SET fulcrum.current_tenant = 'tenant-uuid';
SELECT * FROM envelopes; -- Returns only tenant's data

Performance Characteristics

Metric Target Achieved
Policy Evaluation <10ms P99 ✅ ~2ms
LLM Analysis <50ms P99 ✅ ~45ms
Event Persistence <5ms P99 ✅ ~3ms
Throughput 10K req/s ✅ Verified
Availability 99.9% Target

Security Architecture

  • Authentication: Clerk with API key fallback
  • Authorization: RBAC with policy-based access
  • Encryption: TLS 1.3 in transit, AES-256 at rest
  • Audit: Complete audit trail with tamper detection
  • Secrets: Environment variables, no hardcoding

Document Version: 1.0
Last Updated: January 6, 2026