Demo

Final act: demonstrate not an AI feature but an operating loop — then peer review, then the credential. As with the agent demo, the hero moment isn't a slick success; it's the system catching a regression. Anyone can show AI answering; you'll show AI being kept trustworthy over time.

The 10-minute walkthrough

1. 1The feature and its risks (1 min): what it does, what 'wrong' costs, the failure mode you most needed to guard.
2. 2The eval suite (2 min): golden-set composition, the checks, the judge calibration number. This is the foundation; show it's real.
3. 3The gate catch, live (2.5 min): open a change that introduces a regression and run the gate on stage. Red build, named failing checks, merge blocked. This is the beat that proves the loop works.
4. 4Observability (2 min): the dashboard, a real trace opened and read, the sampled online-eval score. Show you can see and debug production.
5. 5Cost & release (2 min): the before/after cost table with quality held, and the canary + auto-rollback in action on a bad version.
6. 6One full loop turn (0.5 min): the flywheel sentence — a production failure became a golden case became a guarded regression. The loop, closed.

Questions you'll be asked

• 'Show me your dashboard going green while the feature is broken — can it?' (If quality can slide under green infra, your observability failed the reliability-gap test.)
• 'How do you know your judge is trustworthy?' — the calibration number.
• 'A regression is live right now — walk me through the next five minutes.' — rollback, then trace, then eval case.
• 'What does one request cost, and how do you keep it there?' — the number and the levers and the budget alert.

Peer review & credential

1. 1Review one peer's stack: design doc, the loop hands-on (make their gate catch a regression you write), and rubric scores — eval rigor / observability / versioning & rollback / cost control / CI & release automation (20% each).
2. 2The one-change rule, final time: incorporate your reviewer's strongest finding before submission, logged with its effect.
3. 3Submit: design doc, the working stack, the operational runbook, the demo, your peer review.

Passing earns Certified LLMOps Engineer (credential format EDOVA-OPS-2026-XXXX): you build eval suites and observability, control cost and latency, gate regressions in CI, and operate AI features in production through the full evaluate→deploy→observe→improve loop. In a field where most people can prompt and few can operate, this is the credential that says you keep AI working after launch.