Reference architecture
How our AI systems are built.
A layered design that keeps agents, workflows, services and data cleanly separated, so each part can improve without breaking the rest.
The stack
Tools & frameworks, layer by layer
Read it from the top: what people and AI assistants touch, down to the infrastructure that keeps it running.
Interface layer
Experience & agents
Where people and AI assistants meet your systems: dashboards, apps and MCP endpoints that let Claude, ChatGPT or Cursor act on your tools.
- Next.js
- React + Vite
- Flutter
- MCP servers
- Claude · ChatGPT · Cursor
- PDF reports
Model layer
Intelligence
LLM inference, reusable prompt and skill libraries, and AI analysis that turns raw data into drafts, scores and recommendations.
- Claude
- Prompt & skills library
- AI personalization
- AI-SEO analysis
Workflow layer
Orchestration
Event-driven automation that connects your tools, triggers agents and routes work, with schedules, retries and human approval steps.
- n8n
- Webhooks & schedules
- MCP tool servers
- Human approval steps
Service layer
Services & APIs
Purpose-built backends: multi-tenant APIs, background jobs and content services, each with scoped access keys.
- Go
- FastAPI
- Node.js
- Strapi CMS
- REST & MCP endpoints
- Scoped API keys
Data layer
Data
Operational and analytical stores with clear ownership, plus the analytics sources your growth decisions depend on.
- PostgreSQL
- MySQL 8
- Redis
- GA4 · Search Console
Infrastructure layer
Platform & operations
Boring, reliable infrastructure under everything else, so the clever parts stay up.
- Docker Compose
- nginx
- systemd
- Let's Encrypt TLS
- Self-hosted on your VPS or ours
Listed tools are the ones we run in our own platforms. Your architecture may use a different mix.
Design principles
/01
Agents act through scoped tools
AI assistants get narrow, auditable access through MCP servers and per-account API keys, never blanket credentials.
/02
Humans stay in the loop
Workflows include approval steps wherever a wrong action would be costly, and make them easy to review.
/03
Separate thinking from doing
Models generate and analyse; services and workflows execute. That keeps behaviour testable and replaceable.
/04
Data has an owner
Each tenant and client keeps clearly separated data, with the right store (PostgreSQL, MySQL, Redis) for each job.
/05
Own your stack
Self-hosted where it makes sense, standard tooling throughout, so you are never locked in to us or a vendor.
/06
Production first
Containers, TLS, backups and monitoring from day one. If it only works in a demo, it is not done.
Want this architecture applied to your business?
We will review your workflows and map them to these layers.