Learn

Prompt, structure outputs, and validate model responses in an application.

Design a grounded RAG pipeline with chunking, retrieval, and citations.

Connect tools, agent loops, state, and human approval safely.

Reason about evaluation, cost, latency, reliability, and security.

Build gold sets, catch LLM-as-judge traps, and gate changes with offline and online eval.

Use hybrid search, rerankers, query rewrite, and citation checks so retrieval actually helps.

Pack the window: system, tools, memory, retrieved docs, and history. Drop, summarize, or pin.

Decide what belongs in the prompt vs a store. Session, user, and org memory, including forgetting.

Know when prompt, RAG, or tools lose, and when SFT, DPO, or LoRA is worth the data cost.

Handle images, PDFs, and audio as inputs. Ground answers, and know when a pipeline beats a VLM.

Plan vs react, authorize tools, retry, sandbox, and hand off to humans. Multi-agent only when needed.

Route, fall back, cache, stream, and pick models as a cost and quality knob.

Trace prompts, tools, retrieval, tokens, and latency so you can inspect and improve the system.

Handle prompt injection, data exfil, tool policy, secrets, and jailbreaks as product risks.

Design for uncertainty, citations, regenerate, review queues, and trust. Ship interfaces, not notebooks.

Pick an architecture for a messy brief: RAG vs tools vs fine-tune vs workflow, and defend the choice.