Architecture

Swiftward is a modular policy enforcement engine for on-premises deployment. Single binary — runs as one process or scales horizontally as separate components.

Components

Ingestion

HTTP and gRPC APIs (sync and async modes)
Event validation and partitioned queuing
Entity-level ordering guarantees

Workers

Policy evaluation against active version
Transactional state mutations
Decision trace generation
Action execution (webhooks, HITL, notifications)

Control API + UI

Policy management (upload, validate, activate, rollback)
Investigation console with filtering and time-series
Entity state browser with mass edit
HITL review queues with SLA tracking
Event replay console
A/B testing and shadow mode management

LLM Gateway

OpenAI-compatible proxy
Real-time PII detection and redaction
Prompt injection detection (ML-based guardrails)
Policy evaluation on every request/response

MCP Gateway

Tool call authorization for AI agents
Policy-based access control for MCP tool execution

Deployment Modes

Mode	Description	Use Case
Single Process	All components in one binary	Development, low-volume production
Role-Based	Separate processes per component	Horizontal scaling, isolation

Storage

PostgreSQL (required)

Single dependency for event queuing, state, policy storage, and audit trail
ACID transactions for state consistency
Familiar ops model (backups, replication, monitoring)
No distributed coordination complexity

Optional Adapters

Adapter	Purpose	When to Use
Kafka	Ingestion buffering, cross-DC replication	>10k events/sec, multi-region
Redis	Signal caching, rate limiting	High cache hit rates, rate limiting
ClickHouse	Long-term analytics and dashboards	Retention beyond Postgres, heavy aggregations

Data Flow

Event arrives (HTTP/gRPC)
Validated and queued (partitioned by entity)
Worker evaluates against active policy
State changes committed atomically
Decision trace written (immutable audit record)
Actions executed + response returned

Key Properties

Deterministic — same event + same state + same policy = same verdict
Replayable — re-evaluate historical events against any policy version
Idempotent — each event processed exactly once
ACID — state mutations in a single transaction
Two-phase execution — pure evaluation first, then commit + side effects
Enterprise logging — syslog integration for existing infrastructure

What Swiftward is NOT

Swiftward is purpose-built for policy enforcement. It is not:

A general BPM/workflow orchestrator — no long-running workflows, human tasks, or complex state machines
A real-time streaming processor — not a replacement for Flink/Spark; processes events individually
A data pipeline — does not transform or route data between systems
An ML platform — can call external ML/LLM services but does not train or host models

If you need general workflow orchestration, consider Temporal, Airflow, or similar. Swiftward focuses on fast, deterministic, auditable policy decisions.