No model until the boxes exist. Architecture review for AI features starts with data flow, trust boundaries, and SLOs on a whiteboard — not with a Bedrock model id in a ticket.
⚡ TL;DR: Draw producers → stores → gateway → model → tools → users. Mark trust boundaries and PII. Write SLOs before prompt drafts. Reject designs that skip the diagram.
Box diagram checklist
[User] → [Product API] → [AI Gateway] → [Model]
│ │
│ ├→ [RAG index]
│ └→ [Tools / twin]
└→ [Audit log]
| Question | Fail if unanswered |
|---|---|
| Where does PII flow? | — |
| What can tools mutate? | — |
| What’s the rollback? | — |
| What’s the faithfulness SLO? | — |
| Who owns eval? | — |
## ADR stub
### Context
### Data flow (link diagram)
### Trust boundaries
### SLOs
### Model/tool choices (after the above)
### Risks
❌ Starting the ADR with “We will use Claude Sonnet” before tools and tenants are drawn.
Failure modes
Diagrams that omit the logging pipeline (where prompts land). If logs leave the trust boundary, that is a feature surface. Also omit “batch eval jobs” that share prod indexes — draw them.
Closing checklist
- [ ] Diagram reviewed by security + platform
- [ ] Trust boundaries labeled
- [ ] SLOs drafted (Day 93)
- [ ] Twin/dry-run called out (Day 88)
- [ ] Model choice is last, not first
Series navigation
Day 90: Project: Incident Timeline Summarizer · Day 92: Threat Model an AI Feature
Last updated September 11, 2026
Discover more from CheatCoders
Subscribe to get the latest posts sent to your email.
