Architecture Lab

Environment landscape & ALM flow

flowchart LR
  PERS["Personal<br/>sandbox · default DLP"] --> TEAM["Team<br/>shared maker envs"]
  TEAM --> DEPT["Department<br/>governed pre-prod"]
  DEPT -->|"pipelines · approvals<br/>env variables · connection refs"| PROD["Managed production<br/>managed solutions"]

DLP classification matrix

Operating model

• → Dataverse roles and access model per environment class; application ownership registered — orphaned apps detected and re-assigned or retired.
• → Managed solutions only beyond dev; environment variables and connection references make promotion reproducible.
• → Center of Excellence: telemetry on makers/apps/flows, maturity model for maker segmentation, playbooks for support.
• → Explicit transition criteria from citizen development to professional IT: data sensitivity, user count, availability needs, integration depth.

Use-case lifecycle

flowchart LR
  A["Idea"] --> B["Classification"] --> C["Architecture"] --> D["Risk review"] --> E["Pilot"]
  E --> F["Readiness review"] --> G["Production"] --> H["Monitoring"] --> I["Retirement"]

Operating model components

• → Intake: every AI use case registered with a business owner AND a technical owner — no orphaned experiments.
• → Risk classification decides the path: approved data categories, approved model catalogue, review depth.
• → Approval boards with security review and production-readiness gates; exception management with expiry dates, not permanent waivers.
• → Cost ownership per use case; monitoring and incident response wired before go-live; model retirement as a first-class phase.

Selection matrix (method — values are use-case-specific)

Routing & lifecycle

• → Task classification + data sensitivity decide the candidate set; quality benchmarks per task family, not global leaderboards.
• → Routing with fallback chains and circuit breakers; version pinning with declared migration windows.
• → Deprecation and rollback plans written at adoption time; consumption budgets per workload.

Controls per phase

Telemetry spine

flowchart TD
  REQ["Request"] --> TR["Trace<br/>prompt ver · model ver · tools"]
  TR --> TOK["Tokens · cost"]
  TR --> QUAL["Quality<br/>groundedness · retrieval · tool success"]
  TR --> SAFE["Safety<br/>filter hits · escalations"]
  TOK --> SLO["SLO evaluation"]
  QUAL --> SLO
  SAFE --> SLO
  SLO --> AL["Alerts · rollback thresholds"]

Example scorecard

• → Availability · P95 latency · cost per successful task
• → Grounded answer rate · tool success rate · drift vs baseline
• → Human escalation rate · critical safety incidents (target: zero)
• → Every SLO has an owner, an alert threshold and a rollback trigger.

Model components

• → Architecture principles + reference architectures as the shared language; standards catalogue with an explicit exception process.
• → Design reviews from documents (ADRs), not slides — decisions recorded, revisited, escalated when needed.
• → Reusable components and platform ownership; technical debt tracked with the same seriousness as features.
• → Architecture community across streams; roadmap alignment so local optimisation never beats global coherence.

Design-review checklist (excerpt)

• → Is the decision recorded as an ADR with alternatives and trade-offs?
• → Identity, data classification and audit path defined before build?
• → Cost model and SLOs stated? Rollback path tested, not just written?
• → Can another team operate this without its authors?