Add a new model in 5 minutes

ExaMLOps discovers models by scanning pipelines/models/*.yaml — one YAML file per model is the single source of truth. Phase 2 added a one-command scaffold; Phase 14 extended it to also generate the per-model YAML.

TL;DR

exa scaffold DemoAD --task anomaly_detection --type classification

# verify auto-discovery picked it up
exa pipeline validate

# run the full HPC pipeline against dummy data
exa pipeline run --model DemoAD --dummy

# run the model-specific tests + the registry-integrity guard
.venv/bin/pytest tests/unit/test_demoad.py tests/unit/test_registry_integrity.py -v

Alternatively: exa scaffold CLI or dashboard wizard

Via the exa CLI (no Makefile needed):

exa scaffold DemoAD                                                # defaults: performance_prediction / regression
exa scaffold DemoAD --task anomaly_detection --type classification
exa scaffold DemoAD --force                                        # overwrite existing files

--task and --type are enum-constrained: exa scaffold --help shows valid choices in [a|b|c] format and tab-completes them after running exa --install-completion. Run from the repo root — it invokes tools/scaffold_model.py directly.

Via the dashboard (browser-based, admin only):

1. Log in as admin at http://localhost:18099
2. Go to Models → New Model
3. Fill in the name, task type, and promotion thresholds in the wizard
4. Preview the generated files before confirming
5. Click Create model to write the files to the repo (requires the dashboard container to be started with the repo bind mount — set automatically in docker-compose.yml)

After either method, the same verification steps apply:

exa pipeline validate
exa pipeline run --model DemoAD --dummy
.venv/bin/pytest tests/unit/test_demoad.py tests/unit/test_registry_integrity.py -v

What gets generated

Four files (all under git, none gitignored):

File	Purpose
modelzoo/seanergys_modelzoo/models/tasks/<task>/<name>/<name>_model.py	SeanergysSklearnModel subclass — replace the placeholder estimator with your real algorithm
pipelines/models/<name_lower>.yaml	Full model config — datasets, features, lifecycle thresholds, Prefect schedule, serving aliases, inference schema, SeanerBUS UUID
pipelines/model_configs/<name>_config.py	Transforms-only Python shim — provides MODEL_CLASS, SUPPORTED_DATASETS, resolve_embedding_type, get_transforms (model-bound callables that cannot be expressed in YAML)
tests/unit/test_<name>.py	Four checks: registered, config loads, inference contract, instantiation

A matching __init__.py is created next to the model file so Python treats the new directory as a package — required for the auto-discovery scan.

SeanerBUS UUID: The scaffold auto-generates a seanerbus_uuid field in the model YAML. The SeanerBUS bridge registers a dedicated req/res handler for this model at that UUID on startup. Share the UUID with HPC teams via exa seanerbus list or the dashboard /seanerbus page.

Available options

exa scaffold
  NAME=<PascalCase>          # required (e.g. DemoAD, MyClassifier, FuguTransformer)
  TASK=<task>                # default: performance_prediction
                             # one of: performance_prediction, power_consumption_prediction, anomaly_detection
  TASK_TYPE=<type>           # default: classification
                             # one of: classification, regression
  FORCE=1                    # overwrite existing files

The full set of knobs (estimator class, promotion metric / threshold / direction, etc.) is exposed via python tools/scaffold_model.py --help if you need to override them — the Makefile target uses the defaults from tools/model_template/cookiecutter.json.

After scaffolding

1. Implement the algorithm. Replace the placeholder model_class and any feature handling in <name>_model.py. The scaffold uses RandomForestClassifier (classification) by default; substitute in your real estimator and update _inject_defaults.
2. Tune lifecycle thresholds. Edit pipelines/models/<name_lower>.yaml — set lifecycle[*].threshold values for Staging, Canary, and Production. The scaffold uses 80%/90%/100% of your --promotion-threshold as starting points.
3. Verify the YAML guard. exa pipeline validate runs all YAML integrity checks in under 5 seconds and catches missing fields, invalid lifecycle directions, and shim-YAML dataset mismatches.

That's it — the seven-step HPC training flow, MLflow registration, the promotion gate, and Ray Serve auto-loading all kick in for the new model without further wiring.