ai-engineer

license	Apache-2.0
name	ai-engineer
description	Build production-ready LLM applications, advanced RAG systems, and intelligent agents. Implements vector search, multimodal AI, agent orchestration, and enterprise AI integrations. Use PROACTIVELY for LLM features, chatbots, AI agents, or AI-powered applications.
allowed-tools	Read,Write,Edit,Glob,Grep,Bash,WebFetch,mcp__sequentialthinking__sequentialthinking
metadata	{"category":"AI & Machine Learning","tags":["llm","rag","agents","ai","production","embeddings"],"provenance":{"kind":"first-party","owners":["port-daddy"]},"pairs-with":[{"skill":"agentic-app-architecture","reason":"Decides the app's overall shape (transparency, memory, execution substrate) before this skill builds the RAG/agent internals inside it."},{"skill":"agentic-infrastructure-2026","reason":"Picks the framework/observability/cost-governance stack this skill's RAG pipeline and agents run on top of."},{"skill":"llm-router","reason":"Supplies the model-selection mechanics behind this skill's Model Routing Strategy decision point."},{"skill":"episodic-memory-algorithms","reason":"Supplies the vector-store/retrieval algorithm internals behind this skill's RAG Component Selection decision point."},{"skill":"prompt-engineer","reason":"Designs and audits the individual prompts this skill's LLM app is built from."}],"io-contract":{"kind":"deliverable","consumes":["[Truncated]","[Truncated]"],"produces":["[Truncated]","[Truncated]"]}}

AI Engineer

Expert in building production-ready LLM applications, from simple chatbots to complex multi-agent systems. Specializes in RAG architectures, vector databases, prompt management, and enterprise AI deployments.

Decision Points

RAG Component Selection

Query Type Assessment:
├── Simple FAQ/Knowledge Lookup
│   ├── Document Count < 1000 → Chroma + text-embedding-3-small
│   └── Document Count > 1000 → Pinecone + text-embedding-3-large
├── Technical/Code Documentation  
│   ├── Budget Constrained → bge-large + pgvector
│   └── Performance Critical → voyage-2 + Weaviate
└── Conversational/Multi-turn
    ├── Memory Required → Agent pattern + context management
    └── Stateless → Standard RAG pipeline

Reranking Decision:
├── Precision Critical (legal, medical) → Always use Cohere Rerank
├── Latency < 200ms → Skip reranking, tune retrieval
├── Budget Constrained → Cross-encoder (bge-reranker-large)
└── Default → Cohere Rerank with top-10 → top-3

Database Selection:
├── Existing Postgres → pgvector extension
├── Need Hybrid Search → Weaviate or Qdrant
├── Managed Service → Pinecone
└── Self-hosted/Local → Chroma or Qdrant

Model Routing Strategy

Complexity Assessment:
├── Keywords Only (FAQ) → Claude Haiku
├── Single Document Reference → Claude Sonnet  
├── Multi-document Synthesis → Claude Opus
└── Code Generation → Claude Sonnet with tools

Token Budget Check:
├── < 1K tokens → Any model
├── 1K-4K tokens → Sonnet/GPT-4
├── 4K-32K tokens → Claude Opus
└── > 32K tokens → Chunk and summarize first

Model-selection mechanics (cost/latency tiering across providers) belong to llm-router — this decision point is about matching task complexity to a tier, not the routing implementation itself.

Agent vs RAG Decision

Task Classification:
├── Static Knowledge Query → Pure RAG
├── Need External APIs → Agent with tools
├── Multi-step Reasoning → Agent with planning
├── Real-time Data Required → Agent with live tools
└── Simple Q&A → RAG with fallback to agent

Failure Modes

Semantic Mismatch Cascade

Symptoms: Good retrieval precision but poor answer relevance, users say "close but not quite right" Detection Rule: If semantic similarity > 0.8 but user satisfaction < 60% Root Cause: Query and document embeddings optimized for different semantic spaces Fix: Switch to domain-specific embedding model or implement query expansion with synonyms

Context Window Overflow

Symptoms: Responses become generic, model ignores specific retrieved context, inconsistent answers Detection Rule: If context utilization ratio < 30% and response generality score > 0.7 : Too many irrelevant chunks diluting relevant information : Implement stricter relevance threshold (>0.8) and dynamic context selection