Plan du cours
Architectural Foundations for AI Systems
- Layers, boundaries, and separation of concerns
- Application of ports and adapters to AI components
- Quality attribute scenarios specific to AI features
- Architecture decision records as a governance mechanism
Evolving Existing Applications
- Incremental modernization approaches versus full rewrites
- Positioning AI services alongside existing monolithic architectures
- Utilizing Strangler and anti-corruption layer patterns
- Establishing integration seams for existing APIs and data sources
LLM Integration Architectural Patterns
- The model gateway as an abstraction boundary
- Synchronous, asynchronous, and streaming integration strategies
- Fallback mechanisms, graceful degradation, and provider failover
- Treating prompt and model configuration as version-controlled assets
Contracts, APIs, and Data Exchange
- Structured output schemas and validation techniques
- Idempotency, retry logic, and rate limiting strategies
- Latency budgets and timeout orchestration
- Managing cost per request as an architectural constraint
End-to-End RAG Architecture
- Ingestion, chunking, and embedding pipelines
- Indexing, retrieval, reranking, and generation stages
- Positioning retrieval relative to application services
- Trade-offs and failure modes at each architectural stage
Vector Databases and Retrieval Design
- Comparison of managed versus self-hosted vector stores
- Dense, sparse, and hybrid search methodologies
- Metadata filtering and tenant isolation strategies
- Re-embedding strategies during model transitions
Retrieval Quality, Grounding, and Evaluation
- Constructing evaluation sets from real-world queries
- Metrics for retrieval accuracy and answer quality
- Citation, grounding, and containment of hallucinations
- Feedback loops that inform index updates
Agent Architectures
- The agent loop and tool invocation mechanisms
- Deterministic workflows versus autonomous planners
- Identifying cases where agents are inappropriate
- Defining agent responsibilities and state management
Multi-Agent Systems and Orchestration
- Supervisor and orchestrator topologies
- Handoff protocols and message contracts
- Shared memory and context management strategies
- Framework comparison based on architectural merits
Service Communication and Tool Protocols
- Event-driven and queue-based coordination patterns
- Exposing tools through standard protocols
- Sagas, compensation mechanisms, and long-running transactions
- Observability across multiple service hops
AI Agents in the Development Workflow
- Integrating coding agents into the delivery pipeline
- Spec-driven development and review gates
- Test generation and automated quality checks
- Guardrails that maintain architectural control
Governance and Software Quality
- Design reviews for AI features
- Management of model and prompt changes
- Regression evaluation as a pipeline stage
- Ownership structures, policies, and approval paths
Security, Privacy, and Data Governance
- Prompt injection risks and defense in depth
- Secrets management, access control, and least privilege principles
- Data residency, personal data handling, and retention policies
- Auditability and human-in-the-loop controls
Scalability, Cost, and Observability
- Tracing across agent chains
- Token usage and cost telemetry
- Caching strategies and latency optimization
- Service levels and incident response for AI services
Capstone: Architecture for an AI-Driven Solution
- Context, component, and deployment views
- Designing retrieval and agent orchestration
- Planning for security, governance, and cost management
- Peer review, critique, and migration roadmap development
Pré requis
- Expérience dans la conception et la maintenance de systèmes logiciels en production
- Connaissances opérationnelles des API REST et de l'intégration de services
- Notions de bases de données relationnelles et de modélisation des données
- Exposition basique aux outils ou API basés sur des LLM
Public cible
- Architectes logiciels
- Responsables techniques et ingénieurs seniors
- Managers techniques responsables des décisions architecturales
Nos clients témoignent (1)
J'ai acquis des connaissances sur la bibliothèque Streamlit en Python et je vais certainement essayer de l'utiliser pour améliorer les applications de mon équipe qui sont actuellement développées avec R Shiny.
Michal Maj - XL Catlin Services SE (AXA XL)
Formation - GitHub Copilot for Developers
Traduction automatique