Architecture¶
Plasma's architecture is organized into eight specialized layers — five in production today, three emerging (see the note below). Each has a dedicated purpose and its own message bus, and communicates through explicit relays — behavior emerges from channel topology, not a central coordinator.
flowchart TB
Cg["Cognition<br/><small>Kafka · data bus</small>"] <-->|relay| In
Cv["Conversation<br/><small>Matrix · utterance bus</small>"] <-->|relay| In
Ia["Interaction<br/><small>Grafana · APIs</small>"] <-->|relay| In
Vi["Vision<br/><small>emerging</small>"] -.->|relay| In
Au["Audition<br/><small>emerging</small>"] -.->|relay| In
In(["Integration · NATS<br/><small>the event backbone — ECST</small>"])
St["Stabilization<br/><small>emerging · watches all layers</small>"] -.->|observes| In
In --> Fd["Foundation<br/><small>Kubernetes · runs & observes everything</small>"]Every layer relays through Integration (the NATS event backbone), and every layer runs on Foundation. Solid arrows are in production; dotted are emerging.
| Layer | Purpose | Representative tech | |
|---|---|---|---|
| Fd | Foundation | Infrastructure and observability — public cloud, private cloud, on-prem, air-gapped | Kubernetes, etcd, Ceph, Flannel |
| In | Integration | Central hub for all business data exchange | NATS, APIs, serverless |
| Cg | Cognition | AI/ML, analytics, real-time intelligence processing | Spark, Druid, Trino, Kafka |
| Cv | Conversation | Natural language processing, chatbots, AI interactions | DeepSeek, Llama, Matrix |
| Vi | Vision | Computer vision, image recognition, visual perception | — |
| Au | Audition | Audio processing, speech recognition, acoustic perception | — |
| Ia | Interaction | Dashboards, visualization, monitoring, external communication | Grafana, Graylog, Prometheus |
| St | Stabilization | System health monitoring and performance optimization across all layers | — |
What's established today
Five layers are in production today — Foundation, Integration, Cognition, Conversation, Interaction. Vision, Audition, and Stabilization are emerging: they extend the same model (their own bus, relays in and out) but aren't built yet.
Stabilization in particular is a planned homeostasis layer — a stabilization agent watching system health and triggering skills to react: either notify a human or apply a policy automatically to nudge the system back toward a balanced state. The model is defined; the agent isn't implemented yet.
How the layers relate¶
- Foundation provisions and runs everything; Integration is the bus every other layer exchanges data through.
- Cognition turns data into intelligence; Vision and Audition (emerging) will add visual and acoustic perception.
- Conversation handles natural-language interaction; Interaction surfaces state to humans and external systems.
- Stabilization (planned) will observe the whole stack and act to keep it balanced — see the note above.
Each layer is realized as a set of components composed into packages. See Concepts for that model, and the CLI reference for how to compose and deploy them.