Three years building Powerhouse, from the core engine to its cloud

Every change in Powerhouse is stored as an operation, so many contributors can edit the same documents and every step stays auditable. That needs a fast document engine and a secure API.

How it works

1. Connectthe editor people work in
2. SwitchboardGraphQL API with auth and roles
3. Document engineoperations, queues, cache
4. StoragePostgres via Prisma, Redis

What we built

• Storage and caching for the document engine: Prisma storage, a document cache, and operation queues with in-memory and Redis backends.
• Switchboard, the GraphQL API: authentication and role checks, protected mutations, health checks and structured logging.
• Sign-in with Renown, Powerhouse's decentralised identity, covered by end-to-end tests.

• TypeScript
• GraphQL
• Prisma
• Redis
• Playwright
• Sentry

Powerhouse needed a place to run its customers' apps in the EU, each one isolated, monitored and deployable in minutes, without every team having to learn Kubernetes. We designed and built that platform.

How it works

1. git pushor the vetra CLI
2. GitHub Actionbuilds and publishes
3. ArgoCDsyncs the desired state
4. k3s on Hetznerprovisioned with Terraform
5. Tenantapp, Postgres, secrets, dashboards

What we built

• Lightweight Kubernetes (k3s) clusters on Hetzner, provisioned with Terraform.
• GitOps deployment with ArgoCD and Helm: a merge to Git is a deploy.
• Managed Postgres, a private container registry, persistent storage, automatic TLS, and secrets in OpenBao.
• Full observability with Prometheus, Loki, Tempo and Grafana, plus a proxy that keeps each tenant's monitoring data separate.
• The Vetra cloud console with one-click add-ons, a command-line tool and a GitHub Action, so teams deploy without touching Kubernetes.

• k3s
• Terraform
• ArgoCD
• Helm
• Cilium
• OpenBao
• CloudNativePG
• Harbor
• Grafana stack
• Go

Teams want to search and automate their own documents, but many can't send contracts, plans or recordings to an outside AI provider. So the model had to come to the data.

How it works

1. Your filesPDF, Office, scans, audio, video
2. Doclingparses into typed blocks
3. Blocks & transcriptsstructured, timestamped
4. Open-weight LLMllama.cpp, no public endpoint
5. Search & workflowsPaperless-ngx, automations

Everything above runs inside the customer's own cluster.

What we built

• Ingestion with Docling: documents become typed, queryable blocks; recordings become timestamped transcripts.
• Workflow connectors to Paperless-ngx, so scanned paperwork flows in automatically.
• An open-weight language model served with llama.cpp inside the cluster.
• Packaged as a one-click add-on in the Vetra cloud console.

• Docling
• llama.cpp
• Open-weight LLM
• Paperless-ngx
• TypeScript
• Kubernetes