LLM Concepts

Engineer LLM fine-tuning with clear prompting and RAG boundaries, governed datasets, LoRA and QLoRA, evaluation gates, and production operations.

Design, version, and evaluate LLM prompts for production — system vs user roles, few-shot, chain-of-thought, and the limits that push you toward RAG, tools, and fine-tuning.

Engineer MCP host–client–server integrations: tools/resources/prompts, Streamable HTTP transports, auth, timeouts, and the 2026-07-28 stateless protocol core—without replacing orchestration or guardrails.

Understand how LLMs invoke external tools and APIs—including computer use and browser use—to extend capabilities beyond text generation.

Engineer typed LLM tool intents with schema validation, authorization, timeouts, idempotency, observability, and safe production execution.

Engineer reliable LLM context budgets across instructions, evidence, history, tools, and output—with retrieval, truncation, summarization, and production observability.

Why free-text LLM output fails for machines, constrained decoding vs prompt-only JSON, OpenAI JSON Schema, Pydantic validation, and production patterns with Instructor-style retries.

Decision models answer typed questions (pick one, score, yes/no) with a probability for every option and no autoregressive decoding. How they differ from LLM structured outputs and classifiers, the /v1/systemone contract, OpenAI’s Decisions API (public beta), and production patterns for routing, triage, and agent gating.

Detecting when LLMs fabricate facts — grounding checks, NLI verification, citation validation, and production detection layers.

Why LLMs hallucinate — next-token prediction is not truth-seeking — and how to mitigate with RAG, tools, structured outputs, abstention, detection, and verification.