Deployment
Six paths. One codebase.
Production runs on AWS, Azure, or on-prem Kubernetes. Federal runs on GovCloud or air-gapped clusters. Operational on AWS and Azure today for commercial & enterprise. Operational at IL5 and on JWICS for federal. Same deployment everywhere.
AWS
Best for: Commercial & enterprise — in production today
Infrastructure-as-code supplied
Azure
Best for: Microsoft-aligned customers — in production today
Commercial and GovCloud paths supported
Azure GovCloud (DoD/DISA)
Best for: Federal & DoD — IL5 in production today
Configured for high-assurance environments with private routing
Kubernetes
Best for: Enterprises standardizing on K8s — in production today
Infrastructure-as-code module included
On-prem (air-gapped)
Best for: JWICS, classified, and disconnected — operational today
Container image + self-hosted LLMs + local storage + self-hosted IDP
Private routing
Best for: IL5 / private-endpoint environments
In production today on GovCloud with private routing
Customer install path
- 1Infrastructure-as-code modules stand up the full deployment in one shot — networking, managed databases, message broker, load balancer, identity, secrets, certificates, container registries, and DNS.
- 2CI/CD on main builds container images and deploys to the cluster (no infrastructure changes on every release).
- 3App secrets stay in your cloud secret store, synced into the cluster at deploy time.
- 4Steady-state prod deploy uses customer-specific variables for cluster, domains, identity, certificates, and DNS.
Air-gapped recipe
Use the local storage adapter, an OpenAI-compatible LLM (vLLM / llama.cpp / Ollama), and a self-hosted IDP. No external network calls required at runtime.
Use the local storage driver — no cloud object store requiredSelf-hosted OpenAI-compatible LLM as the providerSelf-hosted IDP for OIDC, or PKI/mTLSMirror images in a private container registrySame deployment, different config — no code changes
Operational footprint (defaults)
Dev cluster
Modern Kubernetes, modest node group sized for dev workloads
Prod cluster
Modern Kubernetes, customer-sized
Application versioning
Image tags pinned per release; reproducible builds
Scheduler scaling
Safe to scale horizontally with database-backed lock acquisition
Live updates
Pub/sub fan-out on per-run channels with safety re-poll
Local dev
Same deployment runs on a laptop; optional docker compose for backing services