Architecture Overview¶

This page summarizes how the core Unison services cooperate to deliver edge-first, real-time generated experiences.

Core Control Plane¶

Intent Graph: First stop for intents coming from the experience renderer, Agent VDI, or I/O services. Normalizes requests and forwards them to the orchestrator.
Orchestrator: Central router and planner. Enforces auth and consent, orchestrates skills/tools, calls inference, and manages conversational or task state.
Policy and Consent: Evaluate safety and consent before actions are executed and provide audit records.
Auth: Issues and validates tokens for service-to-service and person-bound flows.
Inference: Gateway to model providers (local-first; remote providers are optional), invoked by the orchestrator when generation is needed.

Context: The consent-aware profile and session store (typically backed by Postgres in the devstack and platform compose).
Context Graph: Maintains graph-shaped context used for recall, relationships, and cross-signal fusion (devstack uses Neo4j; other deployments may run the graph service with Postgres-backed persistence).
Storage: Durable KV + artifacts + vault + audit, exposed through a single service API (metadata in Postgres; artifacts stored on the local storage volume).

Together, these services implement secure edge profiles that other components can read and update under policy and consent.

Experience Renderer: The real-time renderer that turns system state into an experience, and turns multimodal inputs into intents for the control plane.
Agent VDI: A desktop/browser automation actuator used when an outcome requires interacting with graphical software.
I/O Services (speech, vision, core, BCI, braille, sign, …): Device-side emitters that normalize modality events into the control plane.

Devstack: Docker Compose wiring for local end-to-end runs. Brings up the control plane, inference, renderer, I/O services, and backing services such as Postgres, Redis, and Neo4j.
Base Images: Shared base images used by service Dockerfiles.
Shared Docs and Libraries: Canonical specs and schemas plus shared Python helpers for auth, tracing, HTTP, and envelope validation.

Inference is a service: the orchestrator calls one inference service over HTTP, keeping model runtimes out of individual services.
Model packs: releases ship versioned “model packs” that describe which models to load for a given modality/task and where to put them on disk. See Model Packs.

Next: see Architecture Deep Dive for flows, boundaries, and model execution details.