Custom desk

Model Deployment Desk

A working surface for teams who already trained a model and now need the boring middle: evidence packets, stack maps, and governance artifacts that survive procurement.

Evaluation workflow

  1. 1. Freeze prompt + data snapshot IDs in the registry before offline scoring reruns.
  2. 2. Run slice finders on geography, language, and revenue band—log empty slices explicitly.
  3. 3. Attach reviewer disagreement codes; do not collapse them into a single accuracy figure.
  4. 4. Publish a “metrics we refuse to claim” subsection beside headline numbers.
Isometric illustration of containers feeding an inference endpoint

Deployment stack map

Map every hop from feature store to autoscaling group. Color paths you own versus vendor-managed surfaces. The desk ships a Figma starter plus JSON for importing into your internal diagram tool.

Deployment stack map schematic Features Router Inference Clients Observability lane (trace IDs + model version tags)

Governance and risk checklist

  • · Data processing agreements signed for every external embedding vendor?
  • · Incident commander named for model rollback, not only service rollback?
  • · Customer-visible failure copy reviewed for honesty under Korean advertising norms?
  • · Retention timers set for prompts that include personal identifiers?
Compliance binder tabs beside a risk matrix spreadsheet

Want this desk inside a cohort? Start with MLOps Delivery Track or email hello@cognidev.digital.