🧪 archetype: mlops

Train your own models without $50k surprise GPU bills.

Building with PyTorch, TensorFlow, Ray, MLflow, W&B, or Kubeflow? GreatCTO auto-detects the mlops archetype and ships dataset versioning (DVC / LakeFS), training cost budgets, drift detection, bias / fairness audit, shadow + canary serving, and EU AI Act high-risk gates from day one.

$ npx great-cto init Back to home ↗

What you avoid

The 5 MLOps bugs that silently corrupt prod.

Without GreatCTO

☐Training run not reproducible — can't roll back to last good model
☐GPU run costs $50k — discovered on the cloud bill
☐Feature drift undetected for 6 weeks — accuracy down 12%
☐No fairness audit — disparate impact lawsuit on hiring model
☐No shadow mode — first deploy causes p99 latency 10× regression
☐Silent regression · invisible cost · regulator letter.

With GreatCTO

✓mlops-reviewer pins dataset version + code commit on every run
✓Hard cost cap + checkpoint cadence + early stopping
✓Evidently / WhyLabs drift detector + alerting at PSI > 0.2
✓Fairness audit at promotion · 4/5 rule disparate impact bound
✓Shadow → canary → full · single-command rollback tested
✓Reproducible · cost-bounded · drift-monitored · audit-ready.

Auto-applied gates

Detected: `mlflow` + `torch` + `dvc` →
mlops archetype.

Compliance auto-suggested: eu-ai-act · nist-ai-rmf · iso42001. Specialist agents activated:

01 · mlops-reviewer

Training + serving lifecycle

Dataset lineage (DVC / LakeFS) · cost budget enforcement · MLflow / W&B model registry · drift detection wired · bias audit per protected attribute · shadow → canary serving · EU AI Act risk tier classification.

02 · ai-eval-engineer

Golden-set regression

tests/eval/EVAL-*.md golden scenarios · accuracy + F1 + per-cohort breakdown · cost-per-eval · cross-model comparison.

03 · ai-security-reviewer

Model security

Prompt injection (if LLM fine-tune) · model extraction · membership inference · adversarial robustness · supply chain (pretrained model provenance).

04 · data-platform-reviewer

Training data PII

PII classification on training data · GDPR Art. 17 erasure → retrain trigger · sub-processor docs · cross-border transfer compliance.

Domain pack overlays

Likely to overlay on mlops.

Packs auto-attach when CLI detects pack-specific signals (e.g. twilio in deps → voice-pack). Each pack adds its own reviewer agents + human gates on top of the base archetype pipeline.

+ Clinical AI

FDA GMLP + SaMD classification + EU AI Act medical

+ Drug Discovery

ChEMBL versioning, applicability domain, ALCOA+, SiLA2, IQ/OQ/PQ