AI-Native Engineering and Cloudflare AI
Design useful AI applications with explicit quality, authority, privacy, cost, and recovery boundaries.
Choose models with task-specific evidence
How would you choose a model for quote drafting?
Build RAG with retrieval and authorization boundaries
Design tenant-safe retrieval for internal incident notes.
Treat agents as bounded workflows with authority
How would you safely let an agent repair a failed build?
Control AI latency, caching, rate limits, and fallbacks
Design a safe cache for a support assistant.
Evaluate AI releases as software and as uncertain behavior
What would your AI release gate measure?
Design a tenant-safe runbook assistant
A fictional SaaS company wants an assistant using public docs and tenant-specific incident notes. It must never reveal another tenant's data and must handle missing evidence and provider outages.
Work through the problem
- Draw ingestion, query, authorization, and inference flows.
- Select candidate Workers, storage, Vectorize, Workers AI, and AI Gateway roles; justify omissions.
- Write five evaluation cases including cross-tenant retrieval, missing evidence, stale sources, provider timeout, and a valid answer.
- Define model and prompt versions, cache boundaries, spending limits, and fallback constraints.
- Write a release gate and a human escalation path.
Deliverables
- Architecture and threat-boundary diagram
- Evaluation matrix and acceptance criteria
- Failure policy and cost measurement plan
Self-review rubric
- Authorization enforced outside the model
- Retrieval and generation assessed separately
- No unsupported provider or feature promises
- Clear abstention, fallback, and human review behavior
Compare with the worked approach
Authenticate before retrieval, enforce tenant and document permissions, preserve source versions, and assemble bounded context. Evaluate supported inference options on the same cases. Observe requests with privacy-aware metadata, use scoped caches only where safe, and bound retries and consumption. Block release on isolation failures; use fixtures for control flow and semantic evaluations for answer quality. All business gains remain hypotheses until measured.