What is Quint?
Quint is an AI agent risk scoring platform that intercepts, classifies, scores, and optionally blocks every action an AI agent takes. It combines deterministic graph reasoning with learned GNN models and LLM fallback to produce explainable 1-100 risk scores in under 50ms.Graph Intelligence
4-layer scoring pipeline: intrinsic risk, GNN structural analysis, policy violations, and temporal anomaly detection
API Reference
RESTful event ingestion, score retrieval, policy management, and compliance justifications
Protobuf Schemas
Shared contract definitions for proxy, CLI, risk evaluation, and audit trail
Architecture
System design, proxy interception, entity-level risk assessment, and deployment topology
Core Capabilities
Graph-Native Risk Scoring
Graph-Native Risk Scoring
90 inference rules across 7 compliance frameworks (GDPR, HIPAA, SOC2, PCI-DSS, OWASP, ISO27001, OWASP LLM/Agentic). Forward-chaining engine evaluates in under 1ms. Compliance ontology with 1,948 nodes and 1,075 edges powers graph-structured RAG.
4-Layer Composite Scoring
4-Layer Composite Scoring
Every event is scored through four signal layers:
- L1: Intrinsic action risk (deterministic, verb + sensitivity + scope)
- L2: Structural GNN score (learned behavioral patterns)
- L3: Policy violation score (customer-specific forward-chaining)
- L4: Temporal anomaly modifier (rate, sequence, time-of-day)
LLM Fallback with Graph RAG
LLM Fallback with Graph RAG
When GraphReasoner confidence falls below 0.8, compliance articles and mitigations are retrieved from Memgraph and injected into the LLM prompt for grounded justifications.
Multi-Tier Caching
Multi-Tier Caching
Redis L1 + PostgreSQL L2 event signature caching. Identical events return cached scores instantly.
Tech Stack
| Component | Technology |
|---|---|
| API Server | FastAPI (async), Uvicorn |
| Database | PostgreSQL (async, partitioned) |
| Cache | Redis (L1) + PostgreSQL (L2) |
| Graph Engine | quint-graph (NetworkX in-memory) |
| Graph DB | Memgraph (Bolt protocol, neo4j driver) |
| GNN | PyTorch Geometric (CPU, SubgraphGNN) |
| LLM | Gemini 2.5 Flash (fallback scorer) |
| GPU Training | Modal (Qwen3-8B-AWQ) |
| Deployment | Railway (API), Modal (GPU) |
| Schemas | Protobuf (buf.build) |
| Proxy | Go (MCP gateway) |
Score Bands
| Score | Risk Level | Default Action | Color |
|---|---|---|---|
| 1-10 | None | Allow silently | Green |
| 11-30 | Low | Allow, log | Blue |
| 31-55 | Medium | Flag for review | Yellow |
| 56-80 | High | Block, alert security team | Orange |
| 81-100 | Critical | Block immediately, escalate | Red |