Last-minute cheat sheet.

Core positioning

Treat Gen AI as an enterprise capability, not a single chatbot project. The focus is use-case prioritization, trusted data, secure architecture, the RAG-versus-fine-tuning decision, LLMOps, evaluation, observability, governance, and measurable business value.

2-minute reference architecture

1. Start with use-case intake and risk tiering.
2. Build governed data ingestion into object storage or a data lake.
3. Classify data, scan for PII/PHI, and preserve lineage and access-control metadata.
4. Chunk, embed, and index documents with hybrid search and metadata filters.
5. Retrieve with reranking and document-level authorization.
6. Generate with guardrails and citation-first prompting.
7. Return grounded answers with citations and a stated confidence/limitation.
8. Log trace, prompt version, model version, retrieval set, latency, cost, and safety result.
9. Evaluate with golden datasets and human feedback on a schedule.
10. Deploy with CI/CD, canary release, monitoring, and rollback.

RAG vs. fine-tuning vs. prompting

• Prompting: best for instructions, style, and task framing.
• RAG: best for private, factual, or frequently changing knowledge.
• Fine-tuning: best for consistent behavior, domain language, or output format.
• RLHF/RLAIF: best when enough preference feedback exists to optimize alignment.

LLMOps checklist

• Use-case inventory and risk tier assigned before build.
• Data and prompt governance in place.
• Prompt/model/tool version tracked per release.
• Golden evaluation set gating any change.
• Canary rollout with automatic rollback criteria.
• Cost, latency, and hallucination-rate monitoring in production.