Topology

This is the overview; every other chapter details one part of it, so read it first. The diagram brief at the end collects every component and connection into one figure.

Remy is one platform image running as a set of roles on a trusted Kubernetes cluster, in front of a managed data plane, with app code running in microVMs on a second, untrusted cluster in a second VPC. Cloudflare is the edge for everything a browser touches. Every AI model call, every database query, and every managed capability an app uses is a call back into the platform.

#The Tiers

The edge. Cloudflare Workers terminate every request for a deployed app, the editor, and the public sites. One Worker resolves the hostname to an app and reverse-proxies to the platform; another fronts dev-box previews; others rewrite hosts for the API, WebSocket, git, files, image and video CDNs. Custom domains arrive through Cloudflare for SaaS with TLS issued automatically. Public assets are cached at the edge.

The platform. One container image, started with a --role flag, deployed as separate Kubernetes Deployments on the trusted cluster: the HTTP API, the WebSocket server, the general and bulk worker fleets, the git server, the app-database service, the inbound mail server, the sandbox proxy, the provisioner, and the sandbox orchestrator. The voice worker is its own image. An internet-facing Application Load Balancer with a TLS 1.3 FIPS policy fronts the HTTP, WebSocket, git, and sandbox-proxy roles; a Network Load Balancer fronts SMTP.

The data plane. Aurora PostgreSQL through RDS Proxy as the system of record. Valkey for cache, pub/sub, and coordination. S3 in four regions for app databases, files, git mirrors, build artifacts, and workspace snapshots. Qdrant for vector retrieval, shared and dedicated. A Kinesis Firehose stream into an Object-Locked S3 bucket for the audit log. KMS for every key.

The execution islands. A second VPC and a second EKS cluster where all app code runs. Three node groups, one per island: release sandboxes serving deployed apps, job runners compiling releases, and dev boxes leased to people in the editor. Every sandbox is a Kata microVM. The sandbox orchestrator on the trusted cluster creates and watches these pods across a VPC peering that admits exactly three ports.

The model layer. Every model call leaves through the platform's model service. First-party providers are reached over the internet through the NAT gateways. Amazon Bedrock is reached over a PrivateLink interface endpoint so inference traffic never leaves the AWS network. Tuned models serve from a vLLM pool on a GPU node group in the trusted cluster.

Outbound services. Amazon SES for outbound email with a tenant per app. LiveKit and Telnyx for voice. OAuth providers and third-party APIs, each declared in the counterparty registry with its hosts and data boundary.

Observability. New Relic APM, Sentry, and OpenTelemetry from every role. CloudWatch Container Insights on the untrusted cluster. GuardDuty with EKS runtime monitoring, CloudTrail, VPC flow logs, pgaudit on the database, and S3 access logs on private buckets.

The full inventory of these components, with where each runs, what it talks to, and where it is defined, is in the Component Inventory appendix. Every connection between them, with its protocol and direction, is in the Connection Matrix appendix.