CCA Advanced — Claude Cert Academy

The architect credential, elevated.

The advanced tier of the Claude Certified Architect program. CCA-F asked: can you architect a Claude application? CCA-A asks: can you keep that application alive across multiple regions, multiple teams, regulated data, hard cost ceilings, and quarterly model upgrades? Five modules — capacity & cost engineering, advanced retrieval architecture, production reliability, regulated-data architecture, and migration & upgrade strategy.

What you'll learn

• Engineer Claude workloads against hard cost and latency budgets using prompt caching, batch API, model routing, and request batching
• Design retrieval architectures that scale beyond single-vector-store RAG: hybrid search, re-ranking, graph augmentation, and freshness pipelines
• Operate Claude applications under SLOs with distributed tracing, fallback model chains, circuit breakers, and graceful degradation
• Architect Claude systems for regulated data: PII redaction, encryption boundaries, regional sovereignty, audit retention, and tenant isolation
• Roll model upgrades through A/B, shadow, and canary deployments without breaking downstream contracts

Course content

01. Capacity, Cost & Latency Engineering

• The cost & latency cube: tokens × price × concurrency
• Prompt caching: cache hits, write costs, and cache key design
• Batch API economics and the 24-hour latency budget
• Model routing: when Haiku is the right answer
• Request batching, fan-out fan-in, and parallel tool calls
• TPM / RPM quotas and the math of headroom
• Streaming vs non-streaming: when latency hides cost
• Scenario: cut a $40k/mo Claude bill in half without changing UX

02. Advanced Retrieval & Knowledge Architecture

• When single-vector RAG breaks
• Hybrid retrieval: BM25 + dense, fusion strategies
• Re-rankers: cross-encoders and Claude as a re-ranker
• Query rewriting and decomposition pre-retrieval
• Contextual retrieval: chunk-level context windows
• Knowledge graph augmentation and entity-linked retrieval
• Freshness pipelines: ingestion lag, invalidation, and re-embed costs
• Scenario: a legal-research assistant returning stale citations

03. Production Reliability & Observability

• SLOs for non-deterministic systems: what to measure
• Distributed tracing across LLM call graphs (OpenTelemetry semantic conventions)
• Fallback model chains and the cross-provider escape hatch
• Circuit breakers, retry budgets, and bulkheads for LLM calls
• Cost-attributed dashboards: per-tenant, per-feature, per-team
• Quality monitoring: drift detection on outputs
• Incident response runbooks for LLM-backed services
• Scenario: 3am page — Claude API 5xx rate spike

04. Regulated-Data Architecture & Tenancy

• Tenant isolation patterns: shared, pooled, and silo
• PII redaction at the boundary: pre-call vs post-call
• Encryption boundaries and the "in use" gap
• Data residency and regional Claude deployments
• Audit log retention, integrity, and SIEM export
• Right-to-erasure and embedding lifecycles
• Compliance frameworks (SOC2, HIPAA, GDPR) — what changes for Claude
• Scenario: a HIPAA-bound multi-tenant Claude product gap analysis

05. Model Migration, Versioning & Upgrade Strategy

• Why "just pin the model ID" is not a migration strategy
• Eval harnesses as a pre-rollout gate
• Shadow traffic and dual-call comparison
• Canary rollouts with quality SLI tripwires
• Prompt contract testing: snapshot & schema diffs
• Deprecation timelines and the model-sunsetting calendar
• Capability-based feature flags vs model-ID feature flags
• Scenario: roll a Sonnet 4 → 5 upgrade on 30 internal customer prompts

Continue to Claude Cert Academy