04 / ML DEPLOYMENT INFRASTRUCTURE
ModelForge
A safer path from model to service. A model registry and deployment control plane with immutable artifacts, weighted canaries, promotion, rollback, and runtime boundaries.
RELEASE CONTROL / MODEL LIFECYCLE
Immutable artifactPromote / rollbackStable fallback
The problem
Shipping an ML model is a release-management problem as much as an inference problem. A failed new version needs to preserve the stable service and its verifiable artifacts.
Engineering decisions
- Bound immutable model artifacts to SHA-256 identities and workspace-scoped access.
- Separated control-plane behavior from pluggable Python/Go and PyTorch/ONNX runtime execution.
- Implemented weighted canaries, promotion, rollback, health checks, and stable fallback.
- Verified the containerized deployment lifecycle, including a deliberate regression that aborted a canary while preserving the stable version.
What the evidence shows
The recorded CI run passed 164 tests and exercised promotion, rollback, and failed-canary recovery. The repository includes AWS/Terraform reference configuration; no public ModelForge service is claimed.
ModelForge is a source and deployment-lifecycle project. Public hosting has not been launched.