Most ML architectures assume they will be rebuilt in eighteen months. A quality-inspection line or a case-handling system has to keep running when the vendor behind a component changes its pricing or disappears. So we ask a different question: what does it cost to replace each part when we have to? This is the stack we reach for when that answer matters for a decade. It is a reference, not a mandate.
The layers
Data plane. Production data already sits in systems that work: a process historian, an MES database, a records archive. We build versioned pipelines against those systems rather than migrating everything first. The pipelines are code, and a training set can be rebuilt bit for bit.
Open-weight models in a registry you control. Weights, training configuration, data snapshot and evaluation results, on the client's own infrastructure. If the weights are on your disk, a provider changing terms is an inconvenience. If they are not, it is an outage you cannot fix.
Portable serving. ONNX where the architecture allows, served with ONNX Runtime or Triton on Kubernetes, and Ray Serve when a pipeline needs Python-level composition. Open interfaces, open source, at least one credible replacement each, and the export path to a second runtime exercised in CI.
The evaluation harness first. Versioned test sets, agreed metrics, a runner in CI. Swap the model, the runtime or the hardware, re-run it, compare the numbers. Without it every upgrade is a leap of faith, which means upgrades stop.
Monitoring and evidence. The same harness watches live traffic, with drift detection that pages a person before it degrades a decision. Model documentation, lineage and decision logs come out of the same machinery that ships the model.
The exit-path test
For each layer, what does it cost to replace? Models: a fine-tune. Serving: a deployment. Data plane: a new pipeline. Orchestration: an open standard with many vendors behind it.
If the answer at any layer is "a rewrite", the architecture isn't sovereign. It's just self-hosted lock-in.
Not the fastest path to a demo. The fastest path to a system that is still running, and still yours, in year ten.