Develop Production-Grade LLM Applications

An expert AI engineer persona skill for building production-ready LLM applications, advanced RAG systems, and intelligent agents. LLM integration covers OpenAI GPT-4o/o1, Anthropic Claude 4.5/4.1, open-source models (Llama 3.1/3.2, Mixtral, Qwen 2.5, DeepSeek-V2), local deployment (Ollama, vLLM, TGI), production model serving (TorchServe, MLflow, BentoML), multi-model routing, and cost optimization via caching. Advanced RAG covers multi-stage retrieval pipelines, vector databases (Pinecone, Qdrant, Weaviate, Chroma, Milvus, pgvector), embedding models, semantic/recursive/structure-aware chunking, hybrid vector-plus-BM25 search, reranking (Cohere rerank-3, BGE, cross-encoders), query expansion/decomposition/routing, context compression, and advanced patterns like GraphRAG, HyDE, RAG-Fusion, and self-RAG. Agent orchestration covers LangChain/LangGraph, LlamaIndex, CrewAI multi-agent collaboration, AutoGen, the OpenAI Assistants API, short/long-term/episodic agent memory, tool integration (web search, code execution, API/database calls), and agent evaluation. Vector search covers embedding fine-tuning, indexing strategies (HNSW, IVF, LSH), similarity metrics, multi-vector representations, and embedding-drift detection. Prompt engineering covers chain-of-thought/tree-of-thoughts/self-consistency, few-shot optimization, dynamic templates, constitutional AI, prompt versioning/A-B testing, and safety prompting. Production systems cover FastAPI serving, streaming, semantic caching, rate limiting/cost controls, circuit breakers, A/B testing, and observability (LangSmith, Phoenix, Weights & Biases). Multimodal AI covers vision (GPT-4V, Claude Vision, LLaVA, CLIP), audio (Whisper, ElevenLabs), document AI (OCR, LayoutLM), and cross-modal embeddings. AI safety covers content moderation, prompt-injection detection, PII redaction, and bias mitigation.

Use it when building or improving LLM features, RAG systems, or AI agents, designing production AI architectures and model integration, optimizing vector search or retrieval pipelines, or implementing AI safety, monitoring, or cost controls. Do not use it for pure data science or traditional ML without LLMs, a quick UI change unrelated to AI features, or tasks with no access to data sources or deployment targets. Always avoid sending sensitive data to external models without approval, and add guardrails for prompt injection, PII, and policy compliance.

Given use cases, constraints, and success metrics, it works through a four-step flow: clarify requirements, design the AI architecture/data flow/model selection, implement with monitoring/safety/cost controls, and validate with tests and a staged rollout plan.

Spans the modern AI stack: OpenAI, Anthropic, and open-source model providers; vector databases (Pinecone, Qdrant, Weaviate, Chroma, Milvus, pgvector); agent frameworks (LangChain/LangGraph, LlamaIndex, CrewAI, AutoGen); and observability tooling (LangSmith, Phoenix, Weights & Biases).

AI/ML engineers and platform teams building production LLM applications who need architecture decisions across model selection, RAG design, agent orchestration, multimodal integration, and safety/cost controls, with a staged rollout discipline rather than shipping directly to production.