ai-engineering

AI Engineering Expertise

Load this skill when building or working on AI/ML applications, agentic systems, RAG pipelines, or LLM integrations.

Core Frameworks

• •LangChain v0.3+: LLM orchestration, prompt management, tool integration, chains
• •LangGraph v0.2+: Stateful multi-agent workflows, state machines, checkpointing
• •LLM Providers: OpenAI, Anthropic, Google - with proper fallback strategies
• •Vector Stores: Chroma, Pinecone, Weaviate for semantic search and RAG
• •Embedding Models: OpenAI, Sentence Transformers, Cohere

Agentic Design Patterns

ReAct (Reasoning + Acting)

• •Agent reasons about what to do, takes action, observes result, repeats
• •Use LangGraph StateGraph with typed state (TypedDict), proper node/edge definitions
• •Implement execution count limits to prevent infinite loops
• •Always handle error states explicitly in the graph

Multi-Agent Collaboration

• •Specialized agents with focused capabilities (researcher, writer, reviewer)
• •Orchestrator pattern for routing tasks to appropriate agents
• •Shared state management across agents
• •Clear handoff protocols and error propagation

RAG (Retrieval-Augmented Generation)

• •Document chunking strategies (semantic, recursive, token-based)
• •Embedding selection based on use case and language
• •Retriever tuning (k value, similarity threshold, MMR for diversity)
• •Context window management and prompt construction
• •Hybrid search (keyword + semantic) for better recall

State Management

• •Immutable updates: Return new state objects, never mutate in place
• •Typed state: Use TypedDict or Pydantic models with strict validation
• •Checkpointing: MemorySaver for dev, persistent backends (PostgreSQL, Redis) for production
• •Thread management: Use thread_id for conversation persistence across invocations

Production Patterns

Resilience

• •LLM fallback chains: primary model -> fallback model -> error response
• •Retry with exponential backoff (tenacity or custom) for transient failures
• •Circuit breaker pattern for external service calls
• •Graceful degradation when AI components fail

Observability

• •LangSmith integration for prompt debugging, trace analysis, and optimization
• •Structured logging with correlation IDs for request tracing
• •Token usage and cost tracking per request/session
• •Prometheus metrics: call counts, latency histograms, error rates, active sessions

Performance

• •LLM response caching (InMemoryCache for dev, Redis for production)
• •Batch processing with semaphore-controlled concurrency
• •Async/await throughout the stack for I/O-bound LLM calls
• •Connection pooling for vector stores and databases

Security

• •Input sanitization: Remove injection patterns, enforce length limits
• •Content moderation: Filter harmful content before and after LLM processing
• •Rate limiting per user/session to control costs
• •Secrets management: API keys via environment, never hardcoded

Testing Strategy

• •Unit tests: Mock LLM responses, test individual nodes/tools in isolation
• •Integration tests: Test full workflow execution with controlled inputs
• •Behavior tests: Validate agent decision-making with representative scenarios
• •Use pytest-asyncio for async agent testing
• •Snapshot testing for prompt templates to catch unintended changes

Quality Gates for AI Code

1. •ruff check with strict rules - zero violations
2. •basedpyright --strict - zero type errors
3. •pytest --cov - 90%+ coverage on non-LLM code
4. •bandit - zero high-severity security issues
5. •LangSmith evaluation scores above threshold for key behaviors

When to Use This Skill

• •Building LangChain/LangGraph agents or workflows
• •Implementing RAG pipelines or semantic search
• •Designing multi-agent systems
• •Optimizing LLM usage (caching, batching, fallbacks)
• •Adding observability to AI applications