Start course

Design choices

What's in, what's out, and why.

8 decisions behind the curriculum. Each title is the decision; open it for the reason and where it shows up.

Every core topic gets its own day

Tool calling is day 3, evals day 5, agents and MCP day 6, production day 7 and guardrails day 8. Connecting to company systems gets a week 4 deep dive.

Python end to end, Docker on day 7

Students arrive knowing Python and partners use it for AI, so the whole app is Python: FastAPI, the Anthropic and OpenAI SDKs, Pydantic and pytest. It's live on Vercel from day 1; on day 7 students package it with Docker, so in week 5 it can run in the partner's own cloud.

No starter code: day 1 starts from an empty folder

Like a forward deployed engineer with a new customer: write down the problem, then get the thinnest slice live the same day, so access, keys and deploy problems surface while they're cheap. An instructor reference app, tagged daily, lets anyone behind catch up.

Build on models; no training in the core

Partner companies need interns who ship applications on existing models. Model training belongs in an ML course, and fine-tuning is a weeks 3–6 deep dive if a project needs it.

Every lab adds to the student's own live app

Each day ends with a commit to the deployed capstone, so nothing is taught in the abstract: students leave able to do it, not just having heard of it.

Coding agents allowed; students explain every line

That's how the job is done now. Students submit the prompts they used, and cold-call walkthroughs in the daily recap check that they understand the code.

Evals get a full day and carry into the internships

Measuring whether an AI app works is what separates a demo from a product, and "is it good enough?" is the first question a partner company asks.

One topic a day; the rest waits for a project that needs it

Ten topics in ten days is already fast. Fine-tuning, voice, images and multi-agent systems go on the weeks 3–6 deep-dive menu.