AI Engineering: the judgment to command LLMs, agents & RAG
AI can write the code. Can it judge the code? This track builds the one thing AI can't hand you — the judgment to test, orchestrate and command AI systems: LLM foundations, prompt engineering, agents, RAG, evaluations and the tooling companies actually run.
Because “use an LLM” isn't judgment
In an AI-coding world, an entry-level engineer's primary value is testing, orchestration and judgment — not syntax recall. So we build it deliberately, and make you prove it on real AI-agent work.
Judgment, not syntax
AI writes the code now. The skill that lasts is the judgment to test, orchestrate and command it — so we build that deliberately, across a full track of LLMs, agents, RAG and evals.
Real AI-agent work
Every concept is exercised on real work: the Prompt Lab runs and scores your prompts, and you operate your own agent — Mysty — propose-and-approve, its audit log becoming your portfolio.
Grounded in what breaks
Hallucination, prompt injection, data leakage, runaway cost — you learn the failure modes and the guardrails, tied to real incidents.
Six modules, from tokens to command
A build-first path from how models work to orchestrating governed AI tools your company can rely on — judgment at every step.
LLM foundations
How large language models actually work: tokens, context windows, temperature, cost and latency trade-offs, and where models fail — so you can judge their output instead of trusting it blind.
Prompt engineering
COSTAR and canonical technique — role framing, few-shot, structured output, delimiters, chain-of-thought and anti-hallucination — practised live in the Prompt Lab, where every rewrite is scored.
RAG done right
Retrieval-augmented generation end to end: chunking, embeddings, vector stores, hybrid search, re-ranking, grounding and citations — and the judgment to tell a grounded answer from a confident hallucination.
Agents & tool use
Planning, tool/function calling, memory, multi-step loops and guardrails. Orchestrate agents that do real work safely — and hold the judgment to know when an agent is the wrong tool.
Evals & observability
You can't command what you can't measure: build eval sets, judge outputs, catch regressions, trace tokens and cost, and keep quality from drifting in production.
Company AI tooling
Build the internal tools companies run: retrieval over private docs, gateways and key management, MCP tool servers, and putting a human — with judgment — safely in the loop.
Four things you'll actually command
Every module ends in something that runs, and every decision lands in your auditable record — not a quiz.
Score a prompt
Write a prompt, watch it run against a real model, and get scored on COSTAR with an optimized rewrite and rationale.
Ship a RAG answer
Chunk and embed a corpus, retrieve, ground the answer with citations, then judge whether it actually improved.
Command an agent
Give Mysty tools and a goal, set the guardrails and memory, and command it — propose-and-approve — as it plans and executes, every decision logged.
Prove it with evals
Build an eval set, judge outputs at scale, and catch the regression before your users do.
Learn to command AI, not just prompt it
AI writes the code; Miatz builds the engineer who commands it. Take the DSAT for the AI-engineering track — plus Mysty, your own agent to operate. Admission is selective.