Builder

AI Engineer

Ship production AI: retrieval, agents, evals, and APIs — not just notebooks.

3–4 weeks of evenings · 26 lessons · 10 labs · 1 hireable project

0/26 lessons on this path

Mastery syllabus

1. 1 · How the machine sees language

   Tokens, context, embeddings — the three pictures every builder must say in one sentence.

   • Words are not how AI sees words10 min
   • The desk has a size10 min
   • Neighbours in meaning-space10 min

2. 2 · Retrieval you can defend

   RAG, grounded answers, and a refuse path. This is the hiring skill.

   • Closed book vs open notes12 min
   • RAG you can defend10 min
   • Prompts that hold10 min

3. 3 · Agents and tools

   A loop with a stop rule. Search is cheap. Send and pay wait for a human.

   • A tiny campus agent11 min
   • Tools, not magic10 min
   • An agent is a clerk with a checklist10 min

4. 4 · Evals and honesty

   Catch a fluent lie before a client does. Gold cases. A gate in the build.

   • Measure before you brag11 min
   • Proof over confidence11 min
   • Where AI must not decide9 min

5. 5 · Ship and speak

   Fine-tune vs RAG. A tiny take-home. The 90-second story.

   • Fine-tune vs RAG10 min
   • A campus project that hires14 min
   • Design the AI feature10 min
   • Crush the take-home10 min
   • The full hiring loop14 min

6. 6 · Production craft

   Sampling, chunking, rerank, injection, structured output — how 2026 teams actually ship.

   • Temperature is a dice, not a mood12 min
   • Cut by meaning, not by length12 min
   • Retrieve wide, then rank12 min
   • The user is not the boss of the rule12 min
   • Ask for a form, not a poem12 min

7. 7 · Operate and undo

   Traces, cost caps, the whiteboard, five lines after a miss.

   • A gold thread through the miss12 min
   • Tokens are rent12 min
   • Whiteboard: user, retrieve, gate12 min
   • Five lines after a miss12 min

Simulation labs

• Token river

  See why long prompts cost and forget.

  Open the lab

• RagDesk

  Retrieve, ground, or refuse. Live.

  Open the lab

• Chunk desk

  Cut by meaning, not by length.

  Open the lab

• Agent desk

  Think → tool → check → stop.

  Open the lab

• EvalSuite

  Twenty cases including a fluent lie.

  Open the lab

• TakehomeRun

  Cut scope. Still ship. Write the miss.

  Open the lab

• Sample desk

  Cool dice for facts.

  Open the lab

• Inject desk

  The hidden rule still wins.

  Open the lab

• JSON forge

  A gate cannot read fog.

  Open the lab

• System desk

  User, retrieve, gate.

  Open the lab

The project that hires

Hireable artefact: a RAG over a real handbook

1. Pick one real document set (policy, lab notes, product FAQ). Not Wikipedia.
2. Write 20 eval cases by hand, including empty retrieval (must refuse).
3. Ship retrieve → answer → cite or refuse. Show the miss, not only the win.
4. Add a human flag for anything that would leave the building.
5. Write two resume lines with a verb, object, and number. Then three mock interviews here.

Interview room

• A chatbot invents a citation. Trace the bug from user question to the gate that should have refused.
• When do you fine-tune instead of RAG? Give a case where RAG is the wrong tool.
• Design a 20-case eval: easy win, hard retrieval, empty retrieval, fluent-lie trap.
• Sketch the system: user, retriever, model, human gate. What must never auto-send?

Resume lines that hire

• Built a retrieval pipeline that reduced unsupported answers by measuring groundedness.
• Shipped an API with tracing, cost caps, and a human-review flag.