ai-engineering-gates

Agentic AI architecture — deciding whether to build an agent at all, then designing it. Branches: design a single- or multi-agent system, decide agent vs. workflow, design tools/MCP integration, design agent memory, design planning/orchestration, debug a looping or failing agent, or add human-in-the-loop safety gating. This is an architecture/design skill — use `fastapi-genai` for implementation and `prompt-engineering` for the actual prompt templates.

FastAPI implementation for generative AI services — the code/Python layer beneath `production-rag` and `agentic-ai`. Fires regardless of which LLM provider is involved (OpenAI, Anthropic, a self-hosted model, or any other) — the FastAPI/Python patterns are the same either way. Branches: build or serve a model-calling endpoint, stream a response token-by-token (SSE/WebSocket), define a type-safe request/response contract, persist conversation or usage data, add auth or AI-specific security (rate limiting, prompt-injection guardrails), cache or batch for cost/latency, test a non-deterministic endpoint, or containerize and deploy the service.

Retrieval-Augmented Generation (RAG) architecture — designing, scaling, or hardening a RAG pipeline for production. Branches: architect a new RAG pipeline, choose a vector database, decide a chunking strategy, evaluate retrieval/generation quality, decide RAG vs. fine-tuning vs. long-context, pick an advanced pattern (agentic/graph/multimodal RAG), or debug a RAG system that's hallucinating or retrieving the wrong thing. This is an architecture/design skill — use `fastapi-genai` to implement it.

Prompt engineering — designing, debugging, evaluating, or chaining a prompt, grounded in the Five Principles of Prompting and first-principles reasoning when the task is ambiguous. Branches: write or design a new prompt, debug a prompt that's failing, evaluate or optimize an existing prompt, apply a named technique (few-shot, chain-of-thought, ReAct, tool-use), produce a structured/RAG output format, or build a multi-step prompt chain.