Capstone: Ship & certify
The model card, the serving-and-lifecycle plan, the economics, the rubric, and the Certified Model Adaptation Engineer credential.
The final phase packages the work the way a model that matters gets shipped: documented, priced, operable, and honest about its edges.
- 1Write the model card (the artifact the industry has converged on, and yours in one page): intended task and only that task; base model and adaptation lineage (dataset versions, run configs); evaluation summary with the by-class table; known limitations and failure modes from your error analysis; the escape/routing behavior downstream systems must handle; data provenance and rights basis; and the do-not-use-for list, taken as seriously as the features.
- 2The lifecycle plan: registry entry, the regression gate for any future swap, monitoring (the escape-rate canary + a weekly eval sample), the re-distillation/retrain trigger with its threshold, rollback, and the named owner. A model without a lifecycle plan is technical debt with a checkpoint file.
- 3The economics page: your project's real logged costs, the serving road chosen with its math, the volume sensitivity, and the honest verdict — 'ships and pays back in N months', 'ships for latency/residency reasons at cost parity', or 'does not clear the bar; recommend the prompting solution' (a passing verdict when the numbers say so — this course grades judgment, not enthusiasm).
- 4Present it in 10 minutes: the gap that justified the project, the data story (one slide — it earned it), the three-column table, one failure you fixed with data and one you couldn't, the economics, and the ship recommendation. Reviewers will ask you to defend one dataset decision and one threshold; the experiment log is your friend.
Certification rubric
- The gate & the brief (20%) — honest ceiling, measured gap, exit criteria; or the well-documented decision not to tune.
- Data discipline (30%) — guide, curriculum, pipeline-as-code, contamination check, audit score, versioned manifests. The heaviest weight, on purpose.
- Training & evaluation rigor (30%) — protocol followed, runs logged, eval on the shipping artifact, sealed slice honored, error analysis with causes, side effects named.
- Ship judgment (20%) — model card, lifecycle plan, economics with sensitivity, and a recommendation the evidence actually supports.
Passing earns the Certified Model Adaptation Engineer credential (ID format EDOVA-FT-2026-XXXX, independently verifiable at edova.ai/verify) — attesting the full adaptation loop: knowing when to tune, building data worth learning from, training with discipline, evaluating without mercy, and shipping with a lifecycle.
The engineering spine is now complete under you: LLMOps operationalizes your model's lifecycle plan (its regression gates and monitors are built for exactly the artifact you just shipped); Agentic AI Systems is where distilled specialists get deployed as fast, cheap tools inside larger agent loops (your triage student is precisely the kind of component its architectures crave); and Securing AI Systems covers the adversarial surface a tuned model adds (training-data extraction, the safety behaviors your forgetting audit watched). And when a colleague says 'we should fine-tune a model' — you are now, professionally and permanently, the person who asks to see the prompting ceiling first.