01build
Reproducible build
One command produces the model artifact, container and config from a pinned commit. No snowflake environments.
DeploymentsShipped where
Shipped where
your data lives.
The part that decides whether AI becomes infrastructure or a slide deck: rollout, scaling, monitoring, cost control and a security review you can pass the first time.
04Hawthorn AI Labs
hawthorn@labs:~/deploy
$hawthorn deploy --target on-prem --canary 5%
$eval suite passed · 6/6 gates
$artifact signed · sha256:4f9c…e21b
$canary 5% · p95 214ms · error 0.03%
$promoting to 100% in 3 stages
rollout healthy
- Targets
- 6 modes
- Rollback
- < 60s
- Offline capable
- Yes
- IaC
- Terraform / Helm
01 / Targets
One artifact, six places to run it.
Portability is a design decision made in week one, not a migration you discover in month six.
01
Cloud
Managed inference on your AWS, GCP or Azure account. We build the Terraform, you keep the project and the bill.
Explore02
On-premise
Runs inside your data centre on your hardware, with no outbound calls required for inference or evaluation.
Explore03
Hybrid
Sensitive workloads stay private while burst capacity and non-sensitive traffic use managed capacity.
Explore04
Air-gapped
Fully offline delivery: signed artifacts, vendored dependencies, offline model registry and documented install.
Explore05
Edge / embedded
Quantised, distilled models running next to the process that needs them, where latency or connectivity forbids a call.
Explore06
Your platform
Kubernetes, ECS, Nomad or a VM. We fit the target you already operate rather than introducing a new one.
Explore02 / Model CI/CD
Review your pipelineRelease engineering for models.
A model is a build artifact. It gets versioned, tested, canaried and rolled back like one.
02test
Eval as CI
The full evaluation suite runs on every candidate. Failing a gate blocks the merge, exactly like a unit test.
03canary
Canary rollout
Traffic shifts in stages with live comparison against the incumbent. Regressions stop the rollout automatically.
04observe
Observe
Traces, cost, latency, drift and failure taxonomy stream to the dashboard your team already watches.
05iterate
Close the loop
Production failures become labelled evaluation cases, so the same mistake cannot ship twice.
03 / Operating discipline
Numbers you can put in a review.
Targets are set with your platform team before launch and reviewed quarterly against real traffic.
Availability99.5% – 99.9%
Agreed per route, measured externally
p95 latencyroute-specific
Budgeted across retrieval, model and tools
Error budgettracked monthly
Burn rate alerts before customers notice
Model driftweekly eval
Auto-rollback if a rolling window regresses
Cost ceilingper 1k runs
Hard limit with routing to cheaper models
RecoveryRTO < 1h
Last-known-good artifact retained and signed
04 / Security
Built to pass the review.
We treat the security questionnaire as a design input, not a document written at the end.
Signed artifacts
Keys you control
Change-controlled
No data retention
Platforms we deploy to
AWS/GCP/Azure/Kubernetes/Terraform/Helm/Nomad/vLLM/Istio/Prometheus/Grafana/OpenTelemetry/ArgoCD/Vault/
DeploymentsOut of the notebook.
Out of the notebook.
Into the pipeline.
Tell us the target environment and constraints. You get a reference architecture, a sizing estimate and a rollout plan.
Typical first response within 24 hours · NDA-friendly